AI Hardware & Software Stack
What is the main reason systolic arrays in AI accelerators compute matrix products efficiently?
- AEach unit reads data directly from main memory every time
- BThere is only one unit, so the circuit is simple
- CData is passed from unit to neighboring unit and reused, so memory reads are few and many units work simultaneously
Show answer and explanation
Answer: C. Data is passed from unit to neighboring unit and reused, so memory reads are few and many units work simultaneously
In a systolic array, data flows in a regular, pulse-like rhythm between multiply-accumulate units arranged in a grid. A value read once is passed from neighbor to neighbor and used in dozens of operations, so a great deal of computation is done relative to the data read from memory. Only short wires between neighbors are needed, so many units can be packed densely.
- A. That would run out of bandwidth.
- B. Many units are arranged in a grid.
- C. Correct. Data reuse is the key.