Deep Learning Accelerators (NCE)

The Neural Compute Engine (NCE) is a hardware accelerator for Deep Neural Network (DNN) workloads. It features a highly configurable pipeline for maximum support of DNN operations, such as Long Short-Term memory (LSTM) and Local Response Norm (LRN). It also leverages sparsity and low precision for optimal performance.

The Neural Compute Engine is built from up to 6 Neural Compute Tiles, where each tile is a primary unit of computing. Each NCE Tile incorporates its own memory, and DPU and ACT-SHAVE compute resources, and can independently work on a single workload at a given time or be aggregated as a cluster of tiles running the same workload.

The NCE Subsystem features efficient Inter-Tile-Interconnect (ITI) with multicasting and broadcasting capability to allow NCE Tile to efficiently share data between tiles, in case NCE Tiles are aggregated to split workload.

The NCE Subsystem incorporates a DMA engine with a compression unit and broadcasting/multicasting capability for populating multiple NCE Tiles Memory concurrently.

For hardware assisted task synchronization, the NCE Subsystem provides barriers and workload FIFOs. Barriers remove as much software overhead as possible through Interrupt Service Routing (ISR )loops and programming sequences to keep the compute and data-movement pipelines full.