GeForce RTX 4080
Status and notes
One modified 32 GiB card. Needs gpu_unlock.rtx4080_experimental=1 and its exact address. Same matched run as the 10de:2702: FP16/FP32 and BF16/FP32 doubled, cuBLAS GEMM 1.94× at 4096³ and 1.61× at 8192³, both cards capped at 250 W.
Measured throughput
| Mode | Stock | Unlocked | Gain |
|---|---|---|---|
f16/f32 | 107.00 TF/s | 213.70 TF/s | 2.00× |
bf16/f32 | 106.90 TF/s | 213.60 TF/s | 2.00× |
f16/f16 | 214.20 TF/s | 213.90 TF/s | 1.00× |
s8/s32 | 427.10 TOP/s | 425.70 TOP/s | 1.00× |
dp4a | 106.00 TOP/s | 105.80 TOP/s | 1.00× |
ffma | 50.80 TF/s | 50.70 TF/s | 1.00× |
cuBLAS f16/f32 4096³ | 100.70 TF/s | 194.90 TF/s | 1.94× |
cuBLAS f16/f32 8192³ | 102.50 TF/s | 164.90 TF/s | 1.61× |