KNOW GPU AND ASIC

This article helps people understand ASIC and GPU which are two major chips used in the AI world. Broadcom makes special purpose ASIC chips. Nvidia makes general purpose GPU chips.

ASIC

— designs specific algorithms or functions into hardware circuits, such as loss functions, activation functions, normalization processes

— can use reduced number precisions for speedup

— uses SRAM (faster, expensive, low capacity)

— can customize parallel and pipelineed circuits

— good at sparse matrix

— energy efficient

— can become obsolete soon due to new AI architecture and algorithms

— higher cost of designing and printing for each generation and iteration

— good at logic/computes, not for video graphics

— more like DIY projects. Programming API different

— can work at datacenter, edge AI; training and inference

GPU

— flexible, works for any algorithm

— software developers customize precisions

— use HBM memory (on board, fast, large capacity)

— no specialized circuits. 30000 cores for parallelism

— works for both sparse and dense matrices

— higher energy consumption

— lasts longer for a wider range of applications

— high technology barrier to entry

— good at video graphics

— general purpose with common API

— can work at datacenter, edge AI; training and inference

FPGA

— ignore this if you don’t want to learn the details

— logic on the chip can be changed (reprogrammed) after manufacturing

— can be used to prototype the ASIC

— inefficient for real applications

— FPGA is to Python as ASIC is to assemble language