Skip to main content
Vansh Verma
← Lab

silent-fp16

confirmedC++ · Metal / Apple M4Systems & data structuresJuly 11, 2026

the question

How does the Apple M4 GPU accumulate a hand-written fp16 simdgroup_matrix matmul?

what came out

In TRUE fp16 (~5 effective mantissa bits at K=65536), unlike NVIDIA tensor cores which contractually widen to fp32 -- undocumented until now. Declaring a float accumulator is honored, and mx.matmul already uses fp32. Exact oracle plus NEON brackets.

method & receipts

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

→ read the code and re-run it

github.com/v-code01/silent-fp16