| 200H | Graphics performance | Intel® Core™ Ultra 9 processor 285H is up to 1.24X better graphic performance compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | System 1: Intel® Core™ Ultra 9 185H (MTL H45W 6+8) Configuration: Processor: Intel® Core™ Ultra 9 185H (MTL H45W 6+8) PL1=45W, PL2=115W, 16C22T, Turbo frequency up to 5.1 GHz Memory: DDR5-5600 2x32GB, Dual Rank Storage: 512GB Samsung PM9A1 NVMe(SoC Attached) Display Resolution: 1920x1080 for all power KPIs and all performance benchmarks. Operating System: Windows 11 with VBS, Defender, Tamper Protection enabled for power and performance benchmarks. Power Plan: Power plan set to DC Balanced for all power KPIs except for MobileMark 25, MobileMark 30 which is measured in DC BAPCo plan. Power plan set to AC Balanced for all performance benchmarks except SYSmark 30 and CrossMark which are measured in AC BAPCo plan. All power KPIs and performance benchmarks run in Administrator mode. Power Mode: Best Power Efficiency mode for power KPIs in DC. Best Performance mode for performance benchmarks in AC. Graphics Driver: 32.0.101.5533(IA and Resnet50), 32.0.101.6104 (3DMark KPIs) Power Limit 1 Time Window (Tau): 56s Thermal Solution: Air Cooled for all power and performance KPIs BIOS Version: MTLPFWI1.R00.4122.D14.2405211331 (IA, GFX and Resnet50), MTLPFWI1.R00.4122.D21.2408281317 (AI benchmarks) Micro Code Patch: 0x1E (IA, GFX and Resnet50), 0x20 (AI benchmarks) System 2: Intel® Core™ Ultra 9 285H (ARL H45W 6+8) Configuration: Processor: Intel® Core™ Ultra 9 285H (ARL H45W 6+8) PL1=45W, PL2=115W, 16C16T, Turbo frequency up to 5.4 GHz Memory: DDR5-6400 2x32GB, Dual Rank Storage: 512GB Samsung PM9A1 NVMe(SoC Attached) Display Resolution: 1920x1080 for all power KPIs and all performance benchmarks. Operating System: Windows 11 with VBS, Defender, Tamper Protection enabled for power and performance benchmarks. Power Plan: Power plan set to DC Balanced for all power KPIs except for MobileMark 25, MobileMark 30 which is measured in DC BAPCo plan. Power plan set to AC Balanced for all performance benchmarks except SYSmark 30 and CrossMark which are measured in AC BAPCo plan. All power KPIs and performance benchmarks run in Administrator mode. Power Mode: Best Power Efficiency mode for power KPIs in DC. Best Performance mode for performance benchmarks in AC. Graphics Driver: 32.0.101.6104 Power Limit 1 Time Window (Tau): 56s Thermal Solution: Air Cooled for all power and performance KPIs BIOS Version: MTLPFWI1.R00.4404.D82.2410171127 Micro Code Patch: 0x113 | System 1: 3DMark Wild Life Extreme Unlimited Graphics Score: 5784 System 2: 3DMark Wild Life Extreme Unlimited Graphics Score: 7154 | As of Dec 2024 |
| 200H | SPECrate2017_fp_base v1.1.9 ICX 2023.2.3 (n-copy) Estimated | Intel® Core™ Ultra 9 processor 285H is up to 1.23X (est.) faster multi-threaded performance compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: 104.98 System 2 measurement: 129.48 | As of Dec 2024 |
| 200H | SPECrate2017_int_base v1.1.9 ICX 2023.2.3 (1-copy) Estimated | Intel® Core™ Ultra 9 processor 285H is up to 1.15X (est.) faster single-threaded performance compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: 11.55 System 2 measurement: 13.31 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.58X faster CPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H CPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: CPU FP16: 92 System 2 measurement: CPU FP16: 145 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.64X faster GPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: GPU FP16: 515 System 2 measurement: GPU FP16: 847 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.12X faster NPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H NPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: NPU FP16: 338 System 2 measurement: NPU FP16: 378 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.58X faster CPU FP32 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H CPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: CPU FP32: 91 System 2 measurement: CPU FP32: 144 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.19X faster GPU FP32 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: GPU FP32: 286 System 2 measurement: GPU FP32: 339 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.77X faster CPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H CPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: CPU INT8: 250 System 2 measurement: CPU INT8: 443 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.57X faster GPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: GPU INT8: 821 System 2 measurement: GPU INT8: 1292 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 9 processor 285H is up to 1.12X faster NPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 9 processor 185H NPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 measurement: NPU INT8: 620 System 2 measurement: NPU INT8: 694 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285H is up to 1.48X faster CPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 9 processor 185H CPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 1.97 System 2 Measurement: 2.91 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285H is up to 1.82X faster GPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 4.84 System 2 Measurement: 8.8 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285H is up to 1.06X faster NPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 9 processor 185H NPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 7.59 System 2 Measurement: 8.03 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285H is up to 1.32X faster CPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 9 processor 185H CPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 3.01 System 2 Measurement: 3.96 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285H is up to 2.59X faster GPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 8.09 System 2 Measurement: 20.96 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285H is up to 1.09X faster NPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 9 processor 185H NPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 8.81 System 2 Measurement: 9.61 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Procyon AI Image Generation | Intel® Core™ Ultra 9 processor 285H is up to 2.72X faster GPU performance in Stable Diffusion 1.5 FP16 compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 128 System 2 Measurement: 348 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Image Generation | Intel® Core™ Ultra 9 processor 285H is up to 1.56X faster GPU performance in Stable Diffusion 1.5 INT8 compared to Intel® Core™ Ultra 9 processor 185H GPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 1472 System 2 Measurement: 2289 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Image Generation | Intel® Core™ Ultra 9 processor 285H is up to 1.15X faster NPU performance in Stable Diffusion 1.5 INT8 compared to Intel® Core™ Ultra 9 processor 185H NPU | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 704 System 2 Measurement: 811 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | ML Perf Client 0.5.0 | Intel® Core™ Ultra 9 processor 285H is up to 2.78X faster average Time to First Token (in sec) in ML Perf Client compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 1.51 System 2 Measurement: 0.54 | As of Dec 2024 |
| 200H | ML Perf Client 0.5.0 | Intel® Core™ Ultra 9 processor 285H is up to 1.03X faster average 2nd + Token Generation Rate (tokens/sec) in ML Perf Client compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 16.62 System 2 Measurement: 17.08 | As of Dec 2024 |
| 200H | Phi 3.5 mini | Intel® Core™ Ultra 9 processor 285H is up to 1.42X faster in Phi 3.5 mini compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 568 System 2 Measurement: 805 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Mistral 7B | Intel® Core™ Ultra 9 processor 285H is up to 1.6X faster in Mistral 7B compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 471 System 2 Measurement: 754 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Llama 3 (8B) | Intel® Core™ Ultra 9 processor 285H is up to 1.58X faster in Llama 3 (8B) compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 470 System 2 Measurement: 742 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Llama 2 (13B) | Intel® Core™ Ultra 9 processor 285H is up to 1.57X faster in Llama 2 (13B) compared to Intel® Core™ Ultra 9 processor 185H | Intel® Core™ Ultra 9 processor 285H | Same as above | System 1 Measurement: 458 System 2 Measurement: 718 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Graphics performance | Intel® Core™ Ultra 7 processor 265H is up to 1.26X better graphic performance compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | System 1: Intel® Core™ Ultra 7 165H (MTL H28W 6+8) Configuration: Processor: Intel® Core™ Ultra 7 165H (MTL H28W 6+8) PL1=28W, PL2=64W, 16C22T, Turbo frequency up to 5.0 GHz Memory: LPDDR5-7467 2x16GB, Dual Rank Storage: 512GB Samsung PM9A1 NVMe(SoC Attached) Display Resolution: 1920x1080 for all power KPIs and all performance benchmarks. Operating System: Windows 11 with VBS, Defender, Tamper Protection enabled for power and performance benchmarks. Power Plan: Power plan set to DC Balanced for all power KPIs except for MobileMark 25, MobileMark 30 which is measured in DC BAPCo plan. Power plan set to AC Balanced for all performance benchmarks except SYSmark 30 and CrossMark which are measured in AC BAPCo plan. All power KPIs and performance benchmarks run in Administrator mode. Power Mode: Best Power Efficiency mode for power KPIs in DC. Best Performance mode for performance benchmarks in AC. Graphics Driver: 32.0.101.5533 (IA and Resnet50), 32.0.101.6104(3DMark KPIs) Power Limit 1 Time Window (Tau): 28s Thermal Solution: Tc=50c for all IA performance measurements. Tj close to 90c and PL1 Time Window = 1s for all Graphics performance measurements. Tc=25c for Idle and MCS2.0 power measurements. Tc=50c for Microsoft Team Video Conferencing power measurements. Tc=40c for the rest of active power KPIs measurements. BIOS Version: MTLPFWI1.R00.4122.D14.2405211331 (IA, GFX and Resnet50), MTLPFWI1.R00.4122.D21.2408281317 (AI benchmarks) Micro Code Patch: 0x1E (IA, GFX and Resnet50), 0x20 (AI benchmarks) System 2: Intel® Core™ Ultra 7 265H (ARL H28W 6+8) Configuration: Processor: Intel® Core™ Ultra 7 265H (ARL H28W 6+8) PL1=28W, PL2=60W, 16C16T, Turbo frequency up to 5.3 GHz Memory: LPDDR5-8400 2x16GB, Dual Rank Storage: 512GB Samsung PM9A1 NVMe(SoC Attached) Display Resolution: 1920x1080 for all power KPIs and all performance benchmarks. Operating System: Windows 11 with VBS, Defender, Tamper Protection enabled for power and performance benchmarks. Power Plan: Power plan set to DC Balanced for all power KPIs except for MobileMark 25, MobileMark 30 which is measured in DC BAPCo plan. Power plan set to AC Balanced for all performance benchmarks except SYSmark 30 and CrossMark which are measured in AC BAPCo plan. All power KPIs and performance benchmarks run in Administrator mode. Power Mode: Best Power Efficiency mode for power KPIs in DC. Best Performance mode for performance benchmarks in AC. Graphics Driver: 32.0.101.6104 Power Limit 1 Time Window (Tau): 28s Thermal Solution: Tc=50c for all IA performance measurements. Tj close to 90c and PL1 Time Window = 1s for all Graphics performance measurements. Tc=25c for Idle and MCS2.0 power measurements. Tc=50c for Microsoft Team Video Conferencing power measurements. Tc=40c for the rest of active power KPIs measurements. BIOS Version: MTLPFWI1.R00.4404.D82.2410171127 Micro Code Patch: 0x113 | System 1: 3DMark Solar Bay Unlimited Graphics Score: 13463 System 2: 3DMark Solar Bay Unlimited Graphics Score: 16928 | As of Dec 2024 |
| 200H | SPECrate2017_fp_base v1.1.9 ICX 2023.2.3 (n-copy) Estimated | Intel® Core™ Ultra 7 processor 265H is up to 1.23X (est.) faster multi-threaded performance compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: 89.03 System 2 measurement: 109.41 | As of Dec 2024 |
| 200H | SPECrate2017_int_base v1.1.9 ICX 2023.2.3 (1-copy) Estimated | Intel® Core™ Ultra 7 processor 265H is up to 1.15X (est.) faster single-threaded performance compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: 10.84 System 2 measurement: 12.44 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.57X faster CPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H CPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: CPU FP16: 77 System 2 measurement: CPU FP16: 121 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.66X faster GPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: GPU FP16: 498 System 2 measurement: GPU FP16: 826 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.11X faster NPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H NPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: NPU FP16: 336 System 2 measurement: NPU FP16: 374 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.56X faster CPU FP32 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H CPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: CPU FP32: 77 System 2 measurement: CPU FP32: 120 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.22X faster GPU FP32 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: GPU FP32: 276 System 2 measurement: GPU FP32: 337 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.71X faster CPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H CPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: CPU INT8: 212 System 2 measurement: CPU INT8: 363 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.6X faster GPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: GPU INT8: 779 System 2 measurement: GPU INT8: 1249 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265H is up to 1.12X faster NPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165H NPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 measurement: NPU INT8: 614 System 2 measurement: NPU INT8: 688 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265H is up to 1.51X faster CPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165H CPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 1.52 System 2 Measurement: 2.29 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265H is up to 1.93X faster GPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 4.34 System 2 Measurement: 8.38 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265H is up to 1.07X faster NPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165H NPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 7.46 System 2 Measurement: 7.96 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265H is up to 1.29X faster CPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165H CPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 2.36 System 2 Measurement: 3.04 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265H is up to 2.8X faster GPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 7.98 System 2 Measurement: 22.33 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265H is up to 1.11X faster NPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165H NPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 8.7 System 2 Measurement: 9.62 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200H | Procyon AI Image Generation | Intel® Core™ Ultra 7 processor 265H is up to 2.85X faster GPU performance in Stable Diffusion 1.5 FP16 compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 123 System 2 Measurement: 350 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Image Generation | Intel® Core™ Ultra 7 processor 265H is up to 1.51X faster GPU performance in Stable Diffusion 1.5 INT8 compared to Intel® Core™ Ultra 7 processor 165H GPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 1502 System 2 Measurement: 2272 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Procyon AI Image Generation | Intel® Core™ Ultra 7 processor 265H is up to 1.15X faster NPU performance in Stable Diffusion 1.5 INT8 compared to Intel® Core™ Ultra 7 processor 165H NPU | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 701 System 2 Measurement: 809 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | ML Perf Client 0.5.0 | Intel® Core™ Ultra 7 processor 265H is up to 2.98X faster average Time to First Token (in sec) in ML Perf Client compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 1.55 System 2 Measurement: 0.52 | As of Dec 2024 |
| 200H | ML Perf Client 0.5.0 | Intel® Core™ Ultra 7 processor 265H is up to 1.09X faster average 2nd + Token Generation Rate (tokens/sec) in ML Perf Client compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 15.55 System 2 Measurement: 16.88 | As of Dec 2024 |
| 200H | Phi 3.5 mini | Intel® Core™ Ultra 7 processor 265H is up to 1.6X faster in Phi 3.5 mini compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 505 System 2 Measurement: 809 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Mistral 7B | Intel® Core™ Ultra 7 processor 265H is up to 1.78X faster in Mistral 7B compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 424 System 2 Measurement: 755 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Llama 3 (8B) | Intel® Core™ Ultra 7 processor 265H is up to 1.73X faster in Llama 3 (8B) compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 410 System 2 Measurement: 708 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | Llama 2 (13B) | Intel® Core™ Ultra 7 processor 265H is up to 1.76X faster in Llama 2 (13B) compared to Intel® Core™ Ultra 7 processor 165H | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 389 System 2 Measurement: 686 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Graphics performance | Intel® Core™ Ultra 7 processor 265U is up to 1.21X better graphic performance compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | System 1: Intel® Core™ Ultra 7 165U (MTL U15W 2+8) Configuration: Processor: Intel® Core™ Ultra 7 165U (MTL U15W 2+8) PL1=15W, PL2=57W, 12C14T, Turbo frequency up to 4.9 GHz Memory: LPDDR5-7467 2x16GB, Dual Rank Storage: 512GB Samsung PM9A1 NVMe Display Resolution: 1920x1080 for all power KPIs and all performance benchmarks. Operating System: Windows 11 with VBS, Defender, Tamper Protection enabled for power and performance benchmarks. Power Plan: Power plan set to DC Balanced for all power KPIs except for MobileMark 25, MobileMark 30 which is measured in DC BAPCo plan. Power plan set to AC Balanced for all performance benchmarks except SYSmark 30 and CrossMark which are measured in AC BAPCo plan. All power KPIs and performance benchmarks run in Administrator mode. Power Mode: Best Power Efficiency mode for power KPIs in DC. Best Performance mode for performance benchmarks in AC. Graphics Driver: 32.0.101.5533 Power Limit 1 Time Window (Tau): 28s Thermal Solution: Tc=50c for all IA performance measurements. Tj close to 90c and PL1 Time Window = 1s for all Graphics performance measurements. Tc=25c for Idle and MCS2.0 power measurements. Tc=50c for Microsoft Team Video Conferencing power measurements. Tc=40c for the rest of active power KPIs measurements.BIOS Version: MTLPFWI1.R00.4122.D14.2405211331 (IA, GFX and Resnet50), MTLPFWI1.R00.4122.D21.2408281317 (AI benchmarks) Micro Code Patch: 0x1E (IA, GFX and Resnet50), 0x20 (AI benchmarks) System 2: Intel® Core™ Ultra 7 265U (ARL U15W 2+8) Configuration: Processor: Intel® Core™ Ultra 7 265U (ARL U15W 2+8) PL1=15W, PL2=57W, 12C14T, Turbo frequency up to 5.3 GHz Memory: LPDDR5-8400 2x16GB, Dual Rank Storage: 512GB Samsung PM9A1 NVMe(SoC Attached) Display Resolution: 1920x1080 for all power KPIs and all performance benchmarks. Operating System: Windows 11 with VBS, Defender, Tamper Protection enabled for power and performance benchmarks. Power Plan: Power plan set to DC Balanced for all power KPIs except for MobileMark 25, MobileMark 30 which is measured in DC BAPCo plan. Power plan set to AC Balanced for all performance benchmarks except SYSmark 30 and CrossMark which are measured in AC BAPCo plan. All power KPIs and performance benchmarks run in Administrator mode. Power Mode: Best Power Efficiency mode for power KPIs in DC. Best Performance mode for performance benchmarks in AC. Graphics Driver: 32.0.101.6104 Power Limit 1 Time Window (Tau): 28s Thermal Solution: Tc=50c for all IA performance measurements. Tj close to 90c and PL1 Time Window = 1s for all Graphics performance measurements. Tc=25c for Idle and MCS2.0 power measurements. Tc=50c for Microsoft Team Video Conferencing power measurements. Tc=40c for the rest of active power KPIs measurements. BIOS Version: MTLPFWI1.R00.4404.D82.2410171127 Micro Code Patch: 0x113 | System 1: 3DMark Solar Bay Unlimited Graphics Score: 6526 System 2: 3DMark Solar Bay Unlimited Graphics Score: 7910 | As of Dec 2024 |
| 200U | SPECrate2017_fp_base v1.1.9 ICX 2023.2.3 (n-copy) Estimated | Intel® Core™ Ultra 7 processor 265U is up to 1.12X (est.) faster multi-threaded performance compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: 50.9 System 2 measurement: 56.8 | As of Dec 2024 |
| 200U | SPECrate2017_int_base v1.1.9 ICX 2023.2.3 (1-copy) Estimated | Intel® Core™ Ultra 7 processor 265U is up to 1.11X (est.) faster single-threaded performance compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: 19.05 System 2 measurement: 21.21 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.1X faster CPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U CPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: CPU FP16: 51 System 2 measurement: CPU FP16: 56 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.04X faster GPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: GPU FP16: 303 System 2 measurement: GPU FP16: 316 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.07X faster NPU FP16 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U NPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: NPU FP16: 335 System 2 measurement: NPU FP16: 358 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.08X faster CPU FP32 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U CPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: CPU FP32: 51 System 2 measurement: CPU FP32: 55 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.03X faster GPU FP32 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: GPU FP32: 171 System 2 measurement: GPU FP32: 176 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.11X faster CPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U CPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: CPU INT8: 151 System 2 measurement: CPU INT8: 168 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.04X faster GPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: GPU INT8: 468 System 2 measurement: GPU INT8: 488 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Computer Vision | Intel® Core™ Ultra 7 processor 265U is up to 1.07X faster NPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ Ultra 7 processor 165U NPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 measurement: NPU INT8: 611 System 2 measurement: NPU INT8: 656 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265U is up to 1.01X faster CPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165U CPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 0.95 System 2 Measurement: 0.96 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200U | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265U is up to 1.17X faster GPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 2.72 System 2 Measurement: 3.18 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200U | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265U is up to 1.06X faster NPU performance in latency-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165U NPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 6.95 System 2 Measurement: 7.34 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200U | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265U is up to 1.08X faster CPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165U CPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 1.28 System 2 Measurement: 1.38 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200U | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265U is up to 1X faster GPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 4.17 System 2 Measurement: 4.19 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200U | Resnet50 v1.5 | Intel® Core™ Ultra 7 processor 265U is up to 1.06X faster NPU performance in throughput-optimized ResNet50 compared to Intel® Core™ Ultra 7 processor 165U NPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 8.49 System 2 Measurement: 9.01 OpenVINO Version: 2024.3 OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application: Resnet50 v1.5 OpenVINO Benchmark Application Throughput and Latency Optimized measurement using benchmark hints only. Quantized (Int8) Sparse Network used for all measurements. All results estimated for a 180 seconds runtime. eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 1. Latency Optimized - Number Inference Requests: 1 / Batch Size: 1, Submission Type: Sync 2. Throughput Optimized - Number of Inference Requests and batch size is device specific, Submission Type: Async | As of Dec 2024 |
| 200U | Procyon AI Image Generation | Intel® Core™ Ultra 7 processor 265U is up to 1.01X faster GPU performance in Stable Diffusion 1.5 FP16 compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 67 System 2 Measurement: 68 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Image Generation | Intel® Core™ Ultra 7 processor 265U is up to 1.03X faster GPU performance in Stable Diffusion 1.5 INT8 compared to Intel® Core™ Ultra 7 processor 165U GPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 896 System 2 Measurement: 921 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Procyon AI Image Generation | Intel® Core™ Ultra 7 processor 265U is up to 1.08X faster NPU performance in Stable Diffusion 1.5 INT8 compared to Intel® Core™ Ultra 7 processor 165U NPU | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 694 System 2 Measurement: 747 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | ML Perf Client 0.5.0 | Intel® Core™ Ultra 7 processor 265U is up to 1.05X faster average Time to First Token (in sec) in ML Perf Client compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 3.02 System 2 Measurement: 2.87 | As of Dec 2024 |
| 200U | ML Perf Client 0.5.0 | Intel® Core™ Ultra 7 processor 265U is up to 1.06X faster average 2nd + Token Generation Rate (tokens/sec) in ML Perf Client compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 12.84 System 2 Measurement: 13.58 | As of Dec 2024 |
| 200U | Phi 3.5 mini | Intel® Core™ Ultra 7 processor 265U is up to 1.06X faster in Phi 3.5 mini compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 325 System 2 Measurement: 343 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Mistral 7B | Intel® Core™ Ultra 7 processor 265U is up to 1.12X faster in Mistral 7B compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 254 System 2 Measurement: 285 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Llama 3 (8B) | Intel® Core™ Ultra 7 processor 265U is up to 1.11X faster in Llama 3 (8B) compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 246 System 2 Measurement: 272 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200U | Llama 2 (13B) | Intel® Core™ Ultra 7 processor 265U is up to 1.01X faster in Llama 2 (13B) compared to Intel® Core™ Ultra 7 processor 165U | Intel® Core™ Ultra 7 processor 265U | Same as above | System 1 Measurement: 246 System 2 Measurement: 249 Procyon Version: 2.10.1523 OpenVINO version: 2024.5 | As of Dec 2024 |
| 200H | AI performance | Intel® Core™ Ultra 9 285 Processor is up to 3.3X better 1st token latency performance vs. Intel® Ultra 9-185H Processor on Llama-3-8B (INT4). Beam: 1, BS: 1, Input: 1024, output: 128 Intel® Core™ Ultra 9 285 Processor is up to 2.3X better image generation performance vs. Intel® Ultra 9-185H Processor on Stable-Diffusion-v1.5, (INT4). Input: 1024, steps: 20, resolution: 512x512 | Intel® Core™ Ultra 9 processor 285H | System 1: Intel® Core™ Ultra 9 285 (Arrow Lake H 8+16) Configuration: Processor: Intel® Core™ Ultra 9 285 Processor H 8P+16E, 45W, 5.6 GHz Max Turbo Memory: 2x16GB DDR5-6400 Storage: 1024 GB Samsung OS: Windows 11 24H2(OS Build 26100.1457) Workload & version: Llama-3-8B (INT4), Stable-Diffusion-v1.5 (FP16). Procyon AI Computer Vision. BIOS Version: IFWI_ARL_HU_A1A0_PSPP_Release_2024WW41.4.01 Intel® HD Graphics Driver: 32.0.101.6104 System 2: Lenovo Yoga Pro Intel® Core™ Ultra 9 185 (Meteor Lake H 6+8) Configuration: Processor: Intel® Core™ Ultra 9 185 H Processor 6P+8E, 45W, 5.1 GHz Max Turbo Memory: 32GB LPDDR5- 7467 MHz Storage: 1TB SSD OS: Windows 11 26100.712 Workload & version: Llama-3-8B (INT4), Stable-Diffusion-v1.5 (FP16). Procyon AI Computer Vision. BIOS Version: R00.4122.D12 Intel® HD Graphics Driver: 32.0.101.5967 | System 1: Measurement: Llama-3-8B, 1st token Latency: 1010 msec. Stable-Diffusion-V1.5: Image generation time: 5.8 sec. System 2: Measurement: Llama-3-8B, 1st token latency: 3387 msec. Stable-Diffusion-V1.5: Image generation time: 13.7 sec. | As of Nov 2024 |
| 200H | Procyon AI Computer Vision (UL Procyon) | Intel® Core™ Ultra 9 285 Processor is up to 2.2X faster GPU Float16 performance in Procyon AI Computer Vision compared to Intel® Ultra 9-185H Processor | Intel® Core™ Ultra 9 processor 285H | System 1: Intel® Core™ Ultra 9 285 (Arrow Lake H 8+16) Configuration: Processor: Intel® Core™ Ultra 9 285 Processor H 8P+16E, 45W, 5.6 GHz Max Turbo Memory: 2x16GB DDR5-6400 Storage: 1024 GB Samsung OS: Windows 11 24H2(OS Build 26100.1457) Workload & version: Llama-3-8B (INT4), Stable-Diffusion-v1.5 (FP16). Procyon AI Computer Vision. BIOS Version: IFWI_ARL_HU_A1A0_PSPP_Release_2024WW41.4.01 Intel® HD Graphics Driver: 32.0.101.6104 System 2: Lenovo Yoga Pro Intel® Core™ Ultra 9 185 (Meteor Lake H 6+8) Configuration: Processor: Intel® Core™ Ultra 9 185 H Processor 6P+8E, 45W, 5.1 GHz Max Turbo Memory: 32GB LPDDR5- 7467 MHz Storage: 1TB SSD OS: Windows 11 26100.712 Workload & version: Llama-3-8B (INT4), Stable-Diffusion-v1.5 (FP16). Procyon AI Computer Vision. BIOS Version: R00.4122.D12 Intel® HD Graphics Driver: 32.0.101.5967 | System 1 measurement: Procyon GPU Float 16: 847 System 2 measurement: Procyon GPU Float 16: 384 | As of Nov 2024 |
| 200H | Media decode performance | Intel® Core™ Ultra 7 processor 265H is up to 5.8X faster in media performance than NVIDIA Jetson AGX Orin | Intel® Core™ Ultra 7 processor 265H | System 1: Intel® Core™ Ultra 7 265H CPU: 16 cores GPU: Built-in Intel Arc GPU with 8 Xe cores with 4d-systolic NPU: Intel AI Boost Memory: 2x8GB DDR5-6400 Storage: 1024 GB SSD OS: Ubuntu 24.04.1 LTS OS BKC: MTL (24.04.1 LTS w/ Kernel 6.11) Workload & version: OpenVINO 2024.4.0 [From Archives, Linux] & DLStreamer 2024.2.1, Compiler: GCC 13.2.0 CPU Plugin Version: 2024.4.0-16579-c3152d32c9c-releases/2024/4 GPU, NPU Plugin Version: 2024.4.0-16579-c3152d32c9c-releases/2024/4 NPU Driver: Public V1.8.0 OpenCL Compute Runtime Version: 24.17.29377.6 System 2: Jetson Orin AGX-64GB CPU: 12-core Arm Cortex-A78AE v8.2 64-bit CPU 3MB L2 + 6MB L3 @ 2.2Ghz GPU: NVIDIA Ampere architecture with 2048 NVIDIA® CUDA® cores and 64 Tensor cores @ 1.3Ghz Memory: 64GB 256-bit LPDDR5 @ 204.8 GB/s Storage: 64GB eMMC 5.1 OS: Ubuntu 22.04.4 LTS OS BKC: Orin (Ubuntu 20.04.05 LTS (5.10), Ubuntu 22.04.4 LTS (5.15.136-tegra)) Workload & version: Jetpack 5.0.2, Jetpack 6.0 Rev 2 Compiler: GCC 11.4.0 CUDA Version: 11.4.x, 12.2 | System 1 Measurement: 46 streams System 2 Measurement: 8 streams Video Test Cases: HEVC 1080p30 - 2mbps HEVC 4k30 - 10mbps AVC 1080p15 AVC 4k15 | As of Nov 2024 |
| 200H | Video analytics end-to-end AI pipeline performance | Intel® Core™ Ultra 7 processor 265H is up to 3.4X faster than NVIDIA Jetson AGX Orin in video analytics end-to-end AI pipeline performance | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 68 streams System 2 Measurement: 20 streams Test Cases: we use 3 end-to-end (E2E )workloads with different AI complexities Self-checkout Retail WL: Media decode (AVC 1080p15 & 4k15)+ pre-processing + Detection (YoloV5s_416x416 @ 15fps) + Object Tracking + pre-processing + Classification model (EfficientNet-B0 @ 15inf/s) E2E AI pipeline (WL1): Media decode (HEVC 1080p30 & 4k30) + pre-processing + Detection (YoloV5s_640x640 @ 10fps) + Tracking + pre-processing + 2 Classification models (Resnet50 @10 inf/s + Mobilenet-V2 @ 10inf/s) E2E AI pipeline (WL2): Media decode (HEVC 1080p30 & 4k30)+ pre-processing + Detection (YoloV5s_640x640 @ 5fps) + Object Tracking + pre-processing + 2 Classification models (Resnet50 @5 inf/s + Mobilenet-V2 @ 5inf/s) Data was collected at different batch sizes for input videos and max performance was listed for each system. measured KPIs are based on the number of streams and # streams/watt | As of Nov 2024 |
| 200H | Video analytics end-to-end AI pipeline performance | Intel® Core™ Ultra 7 processor 265H is up to 8.2X better performance/Watt/$ than NVIDIA Jetson AGX Orin in video analytics end-to-end AI pipeline performance | Intel® Core™ Ultra 7 processor 265H | Same as above | System 1 Measurement: 68 streams, 37.1 watt, 1.8 stream/watt, 2.3 stream/watt/$ System 1 ASP: $800 System 2 Measurement: 20 streams, 39.9 watt, 0.5 stream/watt, 0.3 stream/watt/$ System 2 ASP: $1799 Test Cases: we use 3 end-to-end (E2E )workloads with different AI complexities Self-checkout Retail WL: Media decode (AVC 1080p15 & 4k15)+ pre-processing + Detection (YoloV5s_416x416 @ 15fps) + Object Tracking + pre-processing + Classification model (EfficientNet-B0 @ 15inf/s) E2E AI pipeline (WL1): Media decode (HEVC 1080p30 & 4k30) + pre-processing + Detection (YoloV5s_640x640 @ 10fps) + Tracking + pre-processing + 2 Classification models (Resnet50 @10 inf/s + Mobilenet-V2 @ 10inf/s) E2E AI pipeline (WL2): Media decode (HEVC 1080p30 & 4k30)+ pre-processing + Detection (YoloV5s_640x640 @ 5fps) + Object Tracking + pre-processing + 2 Classification models (Resnet50 @5 inf/s + Mobilenet-V2 @ 5inf/s) Data was collected at different batch sizes for input videos and max performance was listed for each system. measured KPIs are based on the number of streams, # streams/watt, #streams/watt/$ | As of Nov 2024 |
| 200S | Graphics performance | Intel® Core™ Ultra 9 processor 285 is up to 2.97X better graphic performance vs. Intel® Core™ i9 processor 14900 | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 1845 System 2: Measurement: 621 | As of Oct 2024 |
| 200S | SPECrate2017_fp_base v1.1.9 ICX 2023.2.3 (n-copy) | Intel® Core™ Ultra 9 processor 285 is up to 1.29X (est.) faster multi-threaded performance vs. Intel® Core™ i9 processor 14900 | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 168.68 System 2: Measurement: 130.97 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 2.33X faster GPU Float16 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 GPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: GPU Float 16: 331 GPU Float 32: 189 GPU INT8: 496 System 2 measurement: GPU Float 16: 142 GPU Float 32: 84 GPU INT8: 240 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 2.25X faster GPU Float32 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 GPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: GPU Float 16: 331 GPU Float 32: 189 GPU INT8: 496 System 2 measurement: GPU Float 16: 142 GPU Float 32: 84 GPU INT8: 240 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 2.07X faster GPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 GPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: GPU Float 16: 331 GPU Float 32: 189 GPU INT8: 496 System 2 measurement: GPU Float 16: 142 GPU Float 32: 84 GPU INT8: 240 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 1.78X faster CPU Float16 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: CPU Float 16: 219 CPU Float 32: 216 CPU INT8: 494 System 2 measurement: CPU Float 16: 123 CPU Float 32: 122 CPU INT8: 402 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 1.77X faster CPU Float32 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: CPU Float 16: 219 CPU Float 32: 216 CPU INT8: 494 System 2 measurement: CPU Float 16: 123 CPU Float 32: 122 CPU INT8: 402 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 1.23X faster CPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: CPU Float 16: 219 CPU Float 32: 216 CPU INT8: 494 System 2 measurement: CPU Float 16: 123 CPU Float 32: 122 CPU INT8: 402 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 3.2X faster NPU Float16/Float32 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: NPU Float 16: 393 NPU Float 32: 393 NPU INT8: 731 System 2 measurement: CPU Float 16: 123 CPU Float 32: 122 CPU INT8: 402 | As of Oct 2024 |
| 200S | Procyon AI Computer Vision (UL Procyon v2.8.1305) | Intel® Core™ Ultra 9 processor 285 is up to 1.8X faster NPU INT8 performance in Procyon AI Computer Vision compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1 measurement: NPU Float 16: 393 NPU Float 32: 393 NPU INT8: 731 System 2 measurement: CPU Float 16: 123 CPU Float 32: 122 CPU INT8: 402 | As of Oct 2024 |
| 200S | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285 is up to 3.39x faster CPU performance in latency-optimized ResNet50 compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 4.98 System 2: Measurement: 1.47 AI Benchmark Model - Resnet v1.5 • Resnet v1.5 Resnet v1.5 - Paper • Sourced from Pytorch Hub • OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application • Resnet50 v1.5 OpenVINO v2024.3 Benchmark Application • Throughput and Latency Optimized measurement using benchmark hints only. • Quantized (Int8) Sparse Network used for all measurements. • All results estimated for a 180 seconds runtime. 1. Latency Optimized • Number Inference Requests: 1 / Batch Size: 1 • Submission Type: Sync 2. Throughput Optimized • Number of Inference Requests and batch size is device specific. GPU batch size is set to 32. • Submission Type: Async •eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 | As of Oct 2024 |
| 200S | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285 is up to 2.08x faster GPU performance in latency-optimized ResNet50 compared to Intel® Core™ i9 processor 14900 GPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 3.14 System 2: Measurement: 1.51 AI Benchmark Model - Resnet v1.5 • Resnet v1.5 Resnet v1.5 - Paper • Sourced from Pytorch Hub • OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application • Resnet50 v1.5 OpenVINO v2024.3 Benchmark Application • Throughput and Latency Optimized measurement using benchmark hints only. • Quantized (Int8) Sparse Network used for all measurements. • All results estimated for a 180 seconds runtime. 1. Latency Optimized • Number Inference Requests: 1 / Batch Size: 1 • Submission Type: Sync 2. Throughput Optimized • Number of Inference Requests and batch size is device specific. GPU batch size is set to 32. • Submission Type: Async •eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 | As of Oct 2024 |
| 200S | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285 is up to 5.6x faster NPU performance in latency-optimized ResNet50 compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 8.26 System 2: Measurement: 1.47 AI Benchmark Model - Resnet v1.5 • Resnet v1.5 Resnet v1.5 - Paper • Sourced from Pytorch Hub • OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application • Resnet50 v1.5 OpenVINO v2024.3 Benchmark Application • Throughput and Latency Optimized measurement using benchmark hints only. • Quantized (Int8) Sparse Network used for all measurements. • All results estimated for a 180 seconds runtime. 1. Latency Optimized • Number Inference Requests: 1 / Batch Size: 1 • Submission Type: Sync 2. Throughput Optimized • Number of Inference Requests and batch size is device specific. GPU batch size is set to 32. • Submission Type: Async •eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 | As of Oct 2024 |
| 200S | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285 is up to 1.6x faster CPU performance in throughput-optimized ResNet50 compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 5.59 System 2: Measurement: 3.5 AI Benchmark Model - Resnet v1.5 • Resnet v1.5 Resnet v1.5 - Paper • Sourced from Pytorch Hub • OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application • Resnet50 v1.5 OpenVINO v2024.3 Benchmark Application • Throughput and Latency Optimized measurement using benchmark hints only. • Quantized (Int8) Sparse Network used for all measurements. • All results estimated for a 180 seconds runtime. 1. Latency Optimized • Number Inference Requests: 1 / Batch Size: 1 • Submission Type: Sync 2. Throughput Optimized • Number of Inference Requests and batch size is device specific. GPU batch size is set to 32. • Submission Type: Async •eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 | As of Oct 2024 |
| 200S | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285 is up to 2.48x faster GPU performance in throughput-optimized ResNet50 compared to Intel® Core™ i9 processor 14900 GPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 4.37 System 2: Measurement: 1.76 AI Benchmark Model - Resnet v1.5 • Resnet v1.5 Resnet v1.5 - Paper • Sourced from Pytorch Hub • OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application • Resnet50 v1.5 OpenVINO v2024.3 Benchmark Application • Throughput and Latency Optimized measurement using benchmark hints only. • Quantized (Int8) Sparse Network used for all measurements. • All results estimated for a 180 seconds runtime. 1. Latency Optimized • Number Inference Requests: 1 / Batch Size: 1 • Submission Type: Sync 2. Throughput Optimized • Number of Inference Requests and batch size is device specific. GPU batch size is set to 32. • Submission Type: Async •eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 | As of Oct 2024 |
| 200S | Resnet50 v1.5 | Intel® Core™ Ultra 9 processor 285 is up to 2.8x faster NPU performance in throughput-optimized ResNet50 compared to Intel® Core™ i9 processor 14900 CPU | Intel® Core™ Ultra 9 processor 285 | System 1: Intel® Core™ Ultra 9 processor 285 (Arrow Lake S 8+16) Configuration: Processor: Intel® Core™ Ultra 9 processor 285 / ARL S 8+16 PL1=65W, PL2=182W, 24C24T, Turbo up to 5.6 GHz Memory: 2x32GB DDR5-6400 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.1591) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: MTLSFWI1.R00.4334.D93.2409111548 Micro Code Patch: 0x110 Intel® HD Graphics Driver: 32.0.101.5869 System 2: Intel® Core™ i9 processor 14900 Configuration: Processor: Intel® Core™ i9 processor 14900 / Raptor Lake - S Refresh, 8+16, PL1=65W, PL2= 219W, 24C32T, Turbo up to 5.8GHz Memory: 2x32GB DDR5-5600 2Rx8 Storage: 512GB Samsung PM9A1 NVMe / 512GB Western Display Resolution: 1920x1080 Operating System: Windows 11 (OS Build 26100.712) with VBS, Defender, Tamper Protection enabled. Power Plan: AC/Ultimate Performance for all benchmarks except SYSmark 25, SYSmark 30, and Crossmark which is measured in AC/BAPCo mode. All benchmarks run in Administrator mode. OS: Windows 11 26100.1591 Workload & version: 3DMark Steel Nomad Light Unlimited BIOS Version: RPLSFWI1.R00.5045.A00.2401260733 Micro Code Patch: 0x127 Intel® HD Graphics Driver: 32.0.101.5542 | System 1: Measurement: 9.96 System 2: Measurement: 3.5 AI Benchmark Model - Resnet v1.5 • Resnet v1.5 Resnet v1.5 - Paper • Sourced from Pytorch Hub • OpenVINO toolchain used to quantize to int8 and "sparsify" to 50% weight sparsity. AI Benchmark Test Framework - OpenVINO Benchmark Application • Resnet50 v1.5 OpenVINO v2024.3 Benchmark Application • Throughput and Latency Optimized measurement using benchmark hints only. • Quantized (Int8) Sparse Network used for all measurements. • All results estimated for a 180 seconds runtime. 1. Latency Optimized • Number Inference Requests: 1 / Batch Size: 1 • Submission Type: Sync 2. Throughput Optimized • Number of Inference Requests and batch size is device specific. GPU batch size is set to 32. • Submission Type: Async •eTOPs: Effective Tera Operations Per second. Sensitive to HW efficiency. eTOPS for Resnet50 v1.5 = FPS * 8.2 / 1000 | As of Oct 2024 |