Skip to main content
Vansh Verma
← Lab

bf16mma

mixedC++ · NEON / Apple M4Systems & data structuresJuly 12, 2026

the question

Does the M4's bf16 matrix instruction (NEON BFDOT/BFMMLA) beat a naive bf16-to-fp32 loop?

what came out

~6.8x faster, but the accuracy 'win' is a single-accumulator strawman: a matched-lane 4-accumulator fp32 loop is more accurate than BFDOT. bf16 buys speed, not accuracy, and it isn't bit-deterministic under regrouping. Exact fp64 oracle.

method & receipts

  • Result: mixed
  • Pre-registered — the prediction was written down and committed before the run.

→ read the code and re-run it

github.com/v-code01/bf16mma