Textbook › Part I The Physics of Semiconductors
CHAPTER 15
Design Styles for Accelerators
To close out Part I, let's take stock of what real AI accelerators actually look like. Everything the earlier chapters built up comes into play here.
- SIMD ── one instruction, many pieces of data
- Apply the same operation to a whole row of data at once. It powers the CPU's vector extensions and is the foundation of the GPU. Highly general-purpose and easy to program.
- Systolic array ── lay the units out in a grid and hand values along
- Arithmetic units are arranged in two dimensions, and data flows from neighbor to neighbor as the computation proceeds. A value read once gets reused over and over, so memory accesses drop dramatically. Extremely strong at matrix multiplication. Google's TPU is the best-known design of this kind.
- Dataflow ── whatever has its data runs next
- Rather than following the order of the instructions, an operation executes as soon as the inputs it needs have arrived. This cuts down on control logic.
- In-memory computing ── compute inside the memory
- Use the physical properties of the memory cells themselves to finish a multiply-accumulate at the very moment the data is read out. It is the ultimate form of eliminating data movement altogether, but precision and design difficulty remain open problems.
Comments
Sign in to comment