Systolic Array (TPU-style)
A grid where numbers pulse through like a heartbeat.
How big?A chip about the size of a postage stamp, with a big square grid of multipliers at its heart.
Real: What it looks like.
Keys: arrows rotate · + / − zoom · 0 reset · 1–4 views · S signal · T tour · L labels · Space spin. Models are stylised and built from code: proportions are honest, details are simplified.
What it is
A systolic array is a grid of tiny multiply-add cells. Weights flow in from the top, data from the side. Each cell multiplies, adds to its running total, and passes the numbers to its neighbour. The work moves through the grid in a wave, like blood pumped by a heart (that's what 'systolic' means).
Because numbers hop cell to cell instead of going back to memory each time, the grid does a huge amount of math per trip to memory.
Google's TPU is the best-known example. Several other companies build their own custom AI chips (ASICs) on similar ideas.
Why AI needs it
Neural networks are mostly one operation, matrix multiply, repeated billions of times. A chip built only for that can be more efficient than a general GPU.
Every labeled part
- 1
Matrix unit
The grid. Turn on Signal to see the diagonal wave as work flows through.
- 2
Weights enter
The model's learned numbers stream in from the top, column by column.
- 3
Data enters
Your input (activations) streams in from the left, row by row.
- 4
Vector unit
Handles everything that isn't matrix multiply: activation functions, normalization, adding things up.
- 5
On-chip memory
Holds the data the grid is working on right now.
- 6
Chip-to-chip links
Wire many chips together into a 'pod' that behaves like one giant accelerator.
- 7
HBM
The same stacked memory GPUs use.
Try it · concept lab
Pump numbers through a grid
Step a 3×3 systolic array and watch it multiply two matrices.
Pulse (clock tick)
0 / 7
Each cell multiplies the pair passing through it and adds to its running total. After 7 pulses every cell holds one answer of A × B, and no cell ever went back to memory.
Big idea: Pass numbers neighbour to neighbour instead of fetching them from memory over and over. That is how AI chips do so much math per byte.
Swap it: other ways to do the same job
GPU
More flexible and runs almost any AI code; the systolic grid is more specialised.
Phone NPU
A tiny cousin inside phones and laptops. Runs small models on-device without the cloud.
FPGA
A chip you can rewire after manufacturing. Flexible, but slower and pricier per unit of math.
Talk about it
- Q1
In this grid, numbers pass from cell to cell like a bucket line. What else works like a bucket line?
- Q2
A chip built for one job is faster at it, but bad at new jobs. When is being a specialist worth it?
For grown-ups: there are no right answers here. Ask a question, then ask "why do you think that?" The reasons matter more than the answer.
Printable question sheet (PDF)