814 4x4 BitNet b1.58 Matrix Multiply Accelerator :: Quicker, easier and cheaper to make your own chip!

814 : 4x4 BitNet b1.58 Matrix Multiply Accelerator

Author: siriboi

Description: SPI-controlled 4x4 W1.58A8 matrix multiply accelerator (BitNet b1.58): uint8 activations x ternary {-1,0,+1} weights, multiplier-free PE-parallel MAC

GitHub repository

Open in 3D viewer

Clock: 50000000 Hz

How it works

A 4x4 matrix multiply accelerator controlled via SPI:

A matrix elements are unsigned 8-bit.
B matrix elements are ternary weights encoded in 2 bits (00=0, 01=+1, 10=-1, 11 treated as 0).
The datapath removes multiplication and uses add/sub/skip only.
Two parallel compute lanes (PE=2) process two output elements at once.

Latency for one full 4x4 result is 32 compute cycles ((16 outputs / 2 lanes) * 4 k-steps), not including SPI transfer overhead.

The SPI command byte format is: {R/W[7], SEL[6:5], ROW[4:3], COL[2:1], 0}

SEL=00: Write matrix A element (1 data byte)
SEL=01: Write matrix B element (1 data byte, only bits [1:0] are used as ternary code)
SEL=10: Control/status. Write bit 0 to start. Read returns {6'b0, done, busy}.
SEL=11: Read matrix C element (3 bytes out, MSB first, signed 20-bit two's-complement result in the lower 20 bits)

The accumulator is 20 bits wide. For ternary weights, the dot-product range per element is [-1020, +1020].

How to test

Connect an SPI master (e.g. microcontroller or RP2040 on the TT demo board) to the SPI pins:

Load matrix A by sending write commands with SEL=00 for each element A[row][col].
Load matrix B by sending write commands with SEL=01 for each element B[row][col] using ternary encoding (00=0, 01=+1, 10=-1).
Start computation by writing 0x01 to the control register (SEL=10).
Poll the status register (read SEL=10) until the done bit (bit 1) is set.
Read results from C[row][col] using read commands with SEL=11 (3 bytes per element).

External hardware

SPI master (directly from the RP2040 on the TT demo board, or any external microcontroller).

Input

Output

Bidirectional

spi_sclk

spi_miso

spi_cs_n

busy

spi_mosi

done

Chip location

814 4x4 BitNet b1.58 Matrix Multiply Accelerator