NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL (NVIDIA CX7) Mean Time Error is 13.4x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

Time error represents the instantaneous offset between the SUT and PTP grandmaster (GM) clocks; while software can effectively compensate for small measurement errors, large errors result in significant, measureable offsets between the GM and SUT. Baseline performance was evaluated under a No-Load condition with minimal background traffic to isolate the network adapter hardware and PTP stack. Conversely, a Loaded condition introduced heavy network traffic via iPerf 2.1.5 (pthreads) to simulate congestion and evaluate PTP robustness against packet delay variation using LinuxPTP 4.4. To generate system contention, multiple iPerf flows were executed simultaneously: one flow transmitted UDP packets at the maximum achievable rate from the device under test to the Calnex Grandmaster (GM), while a secondary flow leveraged two ports on another network adapter-isolated in separate network namespaces and connected back-to-back-to drive bidirectional 25 Gbps TCP streams.

Hardware & Network Configuration

Testing was conducted on an Intel Kaseyville motherboard (BIOS KVLDCRB1.KWO.0030.D39.2502281359, built 02/28/2025) powered by an Intel® Xeon® 6556P-B processor (36 CPUs, 2.3 GHz base / 3.5 GHz max, 144M LLC Cache) and 64 GB of dual-channel Hynix DDR5 RDIMM (HMCG88AGBRA190N). High-precision connectivity was evaluated across three distinct 25Gbps network adapters:

Intel® Ethernet Network Adapter E830-XXVDA2 using the ice 2.3.3 driver (firmware 1.00 0x80016fcb 1.3832.0)

NVIDIA MCX713104AS-ADAT ConnectX-7 HHHL using the mlx5_​core driver (firmware 28.37.1014)

Broadcom BCM957504-P425G using the bnxt_​en driver (firmware 226.0.145.1/pkg 226.1.107.1)

Software Environment & OS Tuning

The system ran Ubuntu 22.04 LTS (Kernel 6.9) with boot parameters ro quiet splash split_​lock_​detect=off vt.handoff=7. To ensure optimal real-time performance and minimize jitter during testing, target operating system tunings were applied: CPUs 1 through 3 were isolated using tuna --cpus=1-3 --isolate, and real-time runtime limits were disabled by executing echo -1 > /proc/sys/kernel/sched_​rt_​runtime_​us.

• Broadcom® BCM957504-P425G |25 Gbps |226.0.145.1/pkg 226.1.107.1 |bnxt_​en | Meantime Error is 13.1x higher than Intel

• NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL| 25 Gbps |1.00 0x80016fcb 1.3832.0 |mlx5_​core |Meantime Error is 13.4x higher than Intel

NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL (NVIDIA CX7) System Time Jitter is 1.7x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

Jitter quantifies timing error variability through the standard deviation of measured time errors, with lower values signifying superior clock stability and precision. Baseline synchronization capabilities were established using LinuxPTP 4.4 under a No-Load test with minimal traffic. A Loaded test introduced heavy network traffic using iPerf 2.1.5 (pthreads) to evaluate synchronization robustness against packet delay variation. Multiple iPerf flows were used to induce system contention: one transmitted UDP packets at the maximum achievable rate from the device under test to a Calnex GM, while another generated bidirectional 25 Gbps TCP streams across two ports on a second network adapter configured in separate network namespaces and connected back-to-back.

Hardware & Network Configurations:

Testing was conducted on an Intel Kaseyville motherboard (BIOS KVLDCRB1.KWO.0030.D39.2502281359, built 02/28/2025) powered by an Intel® Xeon® 6556P-B processor (36 CPUs, 2.3 GHz base / 3.5 GHz max, 144M LLC Cache) and 64 GB of dual-channel Hynix DDR5 RDIMM (HMCG88AGBRA190N). High-precision connectivity was evaluated across three distinct 25Gbps network adapters:

Intel® Ethernet Network Adapter E830-XXVDA2 using the ice 2.3.3 driver (firmware 1.00 0x80016fcb 1.3832.0)

NVIDIA MCX713104AS-ADAT ConnectX-7 HHHL using the mlx5_​core driver (firmware 28.37.1014)

Broadcom BCM957504-P425G using the bnxt_​en driver (firmware 226.0.145.1/pkg 226.1.107.1)

Software and OS Tuning:

The system ran Ubuntu 22.04 LTS (Kernel 6.9) with boot parameters ro quiet splash split_​lock_​detect=off vt.handoff=7. To ensure optimal real-time performance and minimize jitter during testing, target operating system tunings were applied: CPUs 1 through 3 were isolated using tuna --cpus=1-3 --isolate, and real-time runtime limits were disabled by executing echo -1 > /proc/sys/kernel/sched_​rt_​runtime_​us.

• Broadcom® BCM957504-P425G |25 Gbps |226.0.145.1/pkg 226.1.107.1 |bnxt_​en | System Time Jitter is 3.6x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

• NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL| 25 Gbps |1.00 0x80016fcb 1.3832.0 |mlx5_​core | System Time Jitter is 1.7x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL (NVIDIA CX7) Maximum Time Error (Peak-To-Peak Error and Outlier) is 5.2x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

Peak-to-Peak Error quantifies the absolute range of observed timing error between the minimum and maximum recorded errors, where smaller values indicate tighter timing control and improved synchronization accuracy. Baseline synchronization capabilities were established using LinuxPTP 4.4 under a No-Load test with minimal traffic. A Loaded test introduced heavy network traffic using iPerf 2.1.5 (pthreads) to evaluate synchronization robustness against packet delay variation. Multiple iPerf flows were used to induce system contention: one transmitted UDP packets at the maximum achievable rate from the device under test to a Calnex GM, while another generated bidirectional 25 Gbps TCP streams across two ports on a second network adapter configured in separate network namespaces and connected back-to-back.

Hardware & Network Configuration:

Testing was conducted on an Intel Kaseyville motherboard (BIOS KVLDCRB1.KWO.0030.D39.2502281359, built 02/28/2025) powered by an Intel® Xeon® 6556P-B processor (36 CPUs, 2.3 GHz base / 3.5 GHz max, 144M LLC Cache) and 64 GB of dual-channel Hynix DDR5 RDIMM (HMCG88AGBRA190N). High-precision connectivity was evaluated across three distinct 25Gbps network adapters:

Intel® Ethernet Network Adapter E830-XXVDA2 using the ice 2.3.3 driver (firmware 1.00 0x80016fcb 1.3832.0)

NVIDIA MCX713104AS-ADAT ConnectX-7 HHHL using the mlx5_​core driver (firmware 28.37.1014)

Broadcom BCM957504-P425G using the bnxt_​en driver (firmware 226.0.145.1/pkg 226.1.107.1)

Software & OS Tuning:

The system ran Ubuntu 22.04 LTS (Kernel 6.9) with boot parameters ro quiet splash split_​lock_​detect=off vt.handoff=7. To ensure optimal real-time performance and minimize jitter during testing, target operating system tunings were applied: CPUs 1 through 3 were isolated using tuna --cpus=1-3 --isolate, and real-time runtime limits were disabled by executing echo -1 > /proc/sys/kernel/sched_​rt_​runtime_​us.

• Broadcom® BCM957504-P425G |25 Gbps |226.0.145.1/pkg 226.1.107.1 |bnxt_​en | Peak-to-Peak Error is 3.7x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

• NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL| 25 Gbps |1.00 0x80016fcb 1.3832.0 |mlx5_​core | Peak-to-Peak Jitter is 5.2x higher than Intel Ethernet Network Adapter E835-XXVDA2 under load

The NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL (NVIDIA CX7) exhibited a 43.4% increase in system time jitter standard deviation compared to its baseline under load, whereas the Intel® Ethernet Network Adapter E830-XXVDA2 maintained a 4.9% change under identical test condition

Degradation under load evaluates a network adapter's precision, efficiency, and stability by measuring jitter under high-bandwidth traffic conditions and comparing the results against an idle-baseline state. Testing compared an idle No-Load state-which isolates inherent adapter hardware and PTP software capabilities under minimal traffic-against a congested Loaded state designed to stress the PTP implementation and measure robustness against packet delay variation. To generate system contention during load testing, iPerf 2.1.5 pthreads drove two distinct flows alongside LinuxPTP 4.4 time synchronization: one transmitting max-rate UDP packets from the device under test to a Calnex GM, and another driving bidirectional 25 Gbps TCP streams across two back-to-back ports in separate network namespaces on a secondary adapter.

Hardware & Network Configuration

Testing was conducted on an Intel Kaseyville motherboard (BIOS KVLDCRB1.KWO.0030.D39.2502281359, built 02/28/2025) powered by an Intel® Xeon® 6556P-B processor (36 CPUs, 2.3 GHz base / 3.5 GHz max, 144M LLC Cache) and 64 GB of dual-channel Hynix DDR5 RDIMM (HMCG88AGBRA190N). High-precision connectivity was evaluated across three distinct 25Gbps network adapters:

Intel® Ethernet Network Adapter E830-XXVDA2 using the ice 2.3.3 driver (firmware 1.00 0x80016fcb 1.3832.0)

NVIDIA MCX713104AS-ADAT ConnectX-7 HHHL using the mlx5_​core driver (firmware 28.37.1014)

Broadcom BCM957504-P425G using the bnxt_​en driver (firmware 226.0.145.1/pkg 226.1.107.1)

Software & OS Tuning:

The system ran Ubuntu 22.04 LTS (Kernel 6.9) with boot parameters ro quiet splash split_​lock_​detect=off vt.handoff=7. To ensure optimal real-time performance and minimize jitter during testing, target operating system tunings were applied: CPUs 1 through 3 were isolated using tuna --cpus=1-3 --isolate, and real-time runtime limits were disabled by executing echo -1 > /proc/sys/kernel/sched_​rt_​runtime_​us.

• Broadcom® BCM957504-P425G |25 Gbps |226.0.145.1/pkg 226.1.107.1 |bnxt_​en | Under the specified test parameters, the system time jitter (standard deviation) increased by 154.7% under load relative to its no-load baseline, compared to a 4.9% increase observed for the Intel® Ethernet Network Adapter E830-XXVDA2.

• NVIDIA® MCX713104AS-ADAT ConnectX®-7 HHHL| 25 Gbps |1.00 0x80016fcb 1.3832.0 |mlx5_​core | Under the specified test parameters, the system time jitter (standard deviation) increased by 43.4% under load relative to its no-load baseline, compared to a 4.9% increase observed for the Intel® Ethernet Network Adapter E830-XXVDA2.

Testing was conducted in April 2026 using Netperf 2.7.0 with custom scripts, including n-perf_​multi.sh for scaling thread counts and n-perfmon_​2Port.py for command execution with statistics monitoring; these scripts were available via a git repository. Each traffic item used eight client threads, and TCP stream/MAERTS tests were run with a 64 KB message size. Power was monitored using Intel® DG Monitor PMC version 3.0.0.2 (internal), while temperature data was collected through the /sys/class/hwmon interface. For IP forwarding validation, driver and firmware configurations were verified against both Rocky Linux 9.7 inbox drivers and the latest vendor releases. The Intel® E835-CQDA2 adapter used NVM version 1.00 (0x00180E0) with the ice driver version 2.5.4 (26R0 PC). The NVIDIA ConnectX-6 DX adapter (CX614106A) used firmware version 22.48.1000 with driver version 26.01-1.0.0. The Broadcom Thor BCM957508-P2100G used firmware version 228.1.111.0 (package 236.1.153.0) with driver version 1.10.3-236.1.155.0. No driver tuning was applied, including CPU affinity adjustments, and the adapter was placed on NUMA node 3 in a PCIe Gen 5 x16 slot. Connectivity was established using QSFP56 1-meter direct attach cables in a back-to-back configuration between ports, with no switch in the data path.

Performance measurements were conducted under bidirectional traffic conditions, achieving approximately 370 Gbps payload throughput per adapter at line rate using eight netperf threads in each direction. Idle power measurements were recorded after the system had been powered on for five minutes, and incremental power was calculated as the difference between loaded and idle conditions. Temperature data was collected using the system hardware monitoring interface. Test by Intel as of April 2026.

• Nvidia ConnectX-6 DX (CX614106A) | 17 Gbps/W | Intel = 1.88× higher

• Broadcom BCM957508-P2100G | 23 Gbps/W | Intel = 1.39× higher

Testing was conducted in April 2026 using Netperf 2.7.0 with custom scripts, including n-perf_​multi.sh for scaling thread counts and n-perfmon_​2Port.py for command execution with statistics monitoring; these scripts were available via a git repository. Each traffic item used eight client threads, and TCP stream/MAERTS tests were run with a 64 KB message size. Power was monitored using Intel® DG Monitor PMC version 3.0.0.2 (internal), while temperature data was collected through the /sys/class/hwmon interface. For IP forwarding validation, driver and firmware configurations were verified against both Rocky Linux 9.7 inbox drivers and the latest vendor releases. The Intel® E835-CQDA2 adapter used NVM version 1.00 (0x00180E0) with the ice driver version 2.5.4 (26R0 PC). The NVIDIA ConnectX-6 DX adapter (CX614106A) used firmware version 22.48.1000 with driver version 26.01-1.0.0. The Broadcom Thor BCM957508-P2100G used firmware version 228.1.111.0 (package 236.1.153.0) with driver version 1.10.3-236.1.155.0. No driver tuning was applied, including CPU affinity adjustments, and the adapter was placed on NUMA node 3 in a PCIe Gen 5 x16 slot. Connectivity was established using QSFP56 1-meter direct attach cables in a back-to-back configuration between ports, with no switch in the data path.

Performance measurements were conducted under bidirectional traffic conditions, achieving approximately 370 Gbps payload throughput per adapter at line rate using eight netperf threads in each direction. Idle power measurements were recorded after the system had been powered on for five minutes, and incremental power was calculated as the difference between loaded and idle conditions. Temperature data was collected using the system hardware monitoring interface. Test by Intel as of April 2026.

• Nvidia ConnectX-6 DX (CX614106A) | 22.15W | Intel = ~47% lower power

• Broadcom BCM957508-P2100G | 16.19W | Intel = ~28% lower power