文献摘自A 2.9TOPS/W Deep Convolutional Neural Network SoC in FD-SOI 28nm for Intelligent Embedded Systems

1 英文缩写

• DCNN: deep convolutional neural networks
• HW: hardware
• DMA: 直接内存访问
• CAF: configurable accelerator framework
• CA: convolution accelerator
• MAC: multiply-accumulate

2 overall architecture

We present a state-of-the-art performance and energy efficient HW-accelerated DCNN processor with the following features:

• An energy-efficient set of DCNN HW convolutional accelerators supporting kernel compression

• an on-chip reconfigurable data-transfer fabric

• a power-efficient array of DSPs to support complete real-world computer vision applications

• an ARM-based host subsystem with peripherals(外围设备)

• a range of high-speed IO interfaces for imaging and other types of sensors

• a chip-to-chip multilink（多链路） to pair multiple devices together
  

• ARM Cortex microcontroller with 128KB of memory

• 不同的外围设备

• 8 DSP clusters

• • 每个DSP cluster有2 DSP, 4-way 16KB instruction caches, 64KB local RAMs and a 64KB shared RAM

• Image&CDNN co-processor subsystem

2 accelerator subsystem

• a configurable fully connected switch to/from different kinds of sources/sinks
• various types of accelerators
  Various kernel sizes (up to 12×12), batch sizes (up to 16), and parallel kernels (up to 4) can be handled by a single CA instance, but any size kernel can be accommodated with the accumulator input.

3 convolutional accelerator

• a line buffer to fetch up to 12 feature map data words in parallel with a single memory access
• 最大kernel size：12*12，但是kernel buffer 有36个读端口
  

4 DSP cluster

32bit DSP
The DSPs are tasked with max or average pooling, nonlinear activation, cross-channel response normalization and classification.