GPUUnlock

BF16 with an FP32 accumulator

Same rate-limit story as FP16/FP32, with BF16's wider exponent. Training and many inference stacks prefer this shape.

bf16/f32 · TF/s · rate-limited by the fuse

1rows
1changed
2.00×peak gain

How this mode was measured

Register-resident mma.m16n16k16 issue rate, 8 independent accumulator chains per warp, best of three timed runs

Distance to the specification peak

  • GeForce RTX 4090: 330.3 TFLOP/s measured against a 330.3 TFLOP/s specification peak — 100.0% of it.

330.3 TFLOP/s (Ada whitepaper, RTX 4090, 128 SMs, 2.52 GHz, FP16 dense) = 1024 FLOP/cycle/SM; measured BF16/FP32 is 189.6–189.7 TF/s vs 189.8–189.9 for FP16, i.e. the same rate (docs/ada-2705-report.md)

Every card measured on this mode

ModeStockUnlockedGainCard
bf16/f32165.50 TFLOP/s330.30 TFLOP/s2.00×GeForce RTX 4090 · 128 SMs

Same card, same driver, same binary. The unlocked figure is the only variable.

Which cards this covers

Only cards with a same-card measurement on this mode appear in the table. A card missing from it has not been measured on this path, which is not the same as failing.

Other measured modes