FP32 FFMA (CUDA cores)
Scalar FP32 fused multiply-add on the CUDA cores, which is the path the unlock never governed.
How this path is measured
Register-resident fma.rn.f32 issue rate: eight independent accumulator chains per warp, best of three timed runs. Library SGEMM rows are grouped with the other cuBLAS GEMMs instead, because a fixed M.N.K call and an issue loop are not the same measurement.
What the unlock changes on this path
8 rows over 8 cards. 0 improved by more than 5%, 8 did not; the largest gain is GeForce RTX 3050 6GB at 1.01×.
| Mode | Stock | Unlocked | Gain | Card |
|---|---|---|---|---|
ffma | 8.40 | 8.50 TF/s | 1.01× | GeForce RTX 3050 6GB |
ffma | 49.4 | 49.7 TF/s | 1.01× | GeForce RTX 4080 SUPER |
ffma | 36.6 | 36.7 TF/s | 1.00× | GeForce RTX 4070 SUPER |
ffma | 40.1 | 40.2 TF/s | 1.00× | GeForce RTX 4070 Ti |
ffma | 79.1 | 79.1 TF/s | 1.00× | GeForce RTX 4090 |
ffma | 31.7 | 31.7 TF/s | 1.00× | GeForce RTX 4070 (AD103) |
ffma | 32.1 | 32.1 TF/s | 1.00× | GeForce RTX 4070 (Gigabyte) |
ffma | 50.8 | 50.7 TF/s | 1.00× | GeForce RTX 4080 |
Reading the numbers
A 1.00× row is not a failed unlock. It says that particular path was never rate-limited, so the override had nothing to restore. If you benchmark your own card and see 1.00× on the mode you care about, the unlock may well have applied correctly and simply does not help that workload.