PYTHON · NUMPY

The MAC Array

Build a multiply-accumulate array that performs a matrix multiply, and count the MACs — the fundamental work of a neural accelerator.

01Challenge

Every neural network is matrix multiplies, and every matrix multiply is a grid of one operation: multiply-accumulate. Build a 4×4 matmul out of MACs, get the exact integer answer, and count how many MAC operations it took.

02Model

Strip an AI accelerator down and you find one circuit repeated thousands of times: the multiply-accumulate, acc += ab. A matrix multiply of an n×k by a k×m matrix is exactly nm*k MACs. A systolic array is just that many MAC units wired in a grid so data flows through and every unit works every cycle. Understanding the MAC count is understanding the chip's job: it is why AI silicon is measured in TOPS, and why moving the data to feed all those MACs is the real design problem.

03Guided practice

Worked: write a single MAC. Faded: triple-loop it into a matmul, counting MACs. Independent: verify against a reference matrix product.

04Feedback

PASS WHEN You built a matrix multiply out of the exact operation an AI accelerator is a grid of — and counted its work in MACs.

If your run fails, check:

FAIL: result wrong — the inner loop accumulates A[i,t]*B[t,j] over t.
FAIL: MAC count off — it should equal n*m*k, one per inner-loop step.

05Retrieve & space

What single operation is an AI accelerator mostly made of? (The multiply-accumulate, acc += a*b.)
How many MACs in an n×k by k×m matmul? (n*m*k.)
What is the hard part of feeding a big MAC array? (Moving data fast enough — the memory/data-movement wall.)

06Mastery & project

Implements a matmul as counted MAC operations and verifies it (L3: the compute primitive of neural hardware).

This integer MAC array is what the next lesson quantizes to int8 to fit on an edge device.

← The Critical Path Quantize for the Edge →

PYTHON · NUMPY

The MAC Array

Do a 4×4 integer matmul with explicit MACs and count them.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you