07 - Deep Learning Processor Architecture ⭐⭐⭐⭐

Architecture design needs to solve two main problems: (1) How to improve the 能效比(performance/power consumption) of the processor - hardening algorithm (2) How to improve the 可编程性(generality) of the processor - CPU

1. Single-core deep learning processor (DLP-S)

1. Overall structure

(1) Architecture diagram

DMAIs a hardware mechanism that allows peripheral components to transfer their I/O data directly into main memory without involving the system processor. 大大提升设备数据传输的吞吐量.

(2) From DLP to DLP-S

Control module : Multiple issue queues, instruction-level parallelism support, register renaming.
Calculation module : ① Increase the operations in the arithmetic unit to support efficient operations performed by hardware; ② Low-bit width arithmetic unit (quantization) to improve execution energy efficiency; ③ Sparse operation to improve calculation