GPUUnlock

FP16/FP32 as a cuBLAS GEMM

A different measurement from the issue-rate rows above: a real cuBLAS GEMM at four M.N.K sizes on Turing. Its 1.36–1.64× must not be averaged with the Ada issue-rate numbers.

f16/f32 · TFLOP/s · rate-limited by the fuse

4rows
4changed
1.64×peak gain

How this mode was measured

cuBLAS GEMM, M.N.K 1024/2048/4096/8192, as tabulated in the nlzy/turing-unlock README

Distance to the specification peak

  • GeForce RTX 2080 Ti: 68.3 TFLOP/s measured against a 56.9 TFLOP/s specification peak — 120.0% of it.
  • GeForce RTX 2080 Ti: 91.4 TFLOP/s measured against a 56.9 TFLOP/s specification peak — 160.6% of it.
  • GeForce RTX 2080 Ti: 81.7 TFLOP/s measured against a 56.9 TFLOP/s specification peak — 143.6% of it.
  • GeForce RTX 2080 Ti: 77.9 TFLOP/s measured against a 56.9 TFLOP/s specification peak — 136.9% of it.

68 SMs × 512 FLOP/cycle/SM × 1.635 GHz = 56.92 TFLOP/s (Turing FP16→FP32 tensor rate)

Every card measured on this mode

ModeStockUnlockedGainCard
f16/f3250.30 TFLOP/s68.30 TFLOP/s1.36×GeForce RTX 2080 Ti · 68 SMs
f16/f3255.70 TFLOP/s91.40 TFLOP/s1.64×GeForce RTX 2080 Ti · 68 SMs
f16/f3253.70 TFLOP/s81.70 TFLOP/s1.52×GeForce RTX 2080 Ti · 68 SMs
f16/f3252.80 TFLOP/s77.90 TFLOP/s1.48×GeForce RTX 2080 Ti · 68 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