B11. Tensor Processing Unit

Year: 2015
Zone: Hardware & Architecture
Band: outer
Tier: Built

AI accelerators are built around matrix multiply. A systolic array pulses data through a grid of multiply-accumulate cells, so every value is reused without another trip to memory.

First TPU in service2015
TPU v1 matrix unit256 × 256
Core operationMultiply-accumulate

Portal

4 × 4 systolic array · cycle 0 of 10 · 0 / 64 MACs

0c00

0c01

0c02

0c03

0c10

0c11

0c12

0c13

0c20

0c21

0c22

0c23

0c30

0c31

0c32

0c33

Orange values of A stream in from the left, white values of B from the top. Each cell multiplies what passes through and adds it to its running total. The lit diagonal is the wavefront.

No cell asks memory for anything mid-computation. Data pulses through the grid, which is why the array is called systolic.

This toy grid has 16 cells. Google's first TPU had a 256 × 256 grid: 65,536 multiply-accumulates every cycle.

Loading portal...

Open in museum View on floor plan All exhibits

Sources

[TBD: source] ([TBD: source])

Portal

Sources

On the map