GeForce RTX 4080 SUPER
状況と注記
One card, measured against its own stock state in the same host as a 10de:2704. Register-resident FP16/FP32 and BF16/FP32 doubled; cuBLAS GEMM reached 1.94× at 4096³ and 1.61× at 8192³ because the large shapes are power-limited at 250 W.
実測スループット
| モード | 純正 | アンロック後 | 向上 |
|---|---|---|---|
f16/f32 | 104.80 TF/s | 209.70 TF/s | 2.00× |
bf16/f32 | 104.80 TF/s | 209.30 TF/s | 2.00× |
f16/f16 | 210.10 TF/s | 210.40 TF/s | 1.00× |
s8/s32 | 418.90 TOP/s | 417.00 TOP/s | 0.99× |
dp4a | 103.90 TOP/s | 103.80 TOP/s | 1.00× |
ffma | 49.40 TF/s | 49.70 TF/s | 1.01× |
cuBLAS f16/f32 4096³ | 98.70 TF/s | 190.70 TF/s | 1.93× |
cuBLAS f16/f32 8192³ | 100.90 TF/s | 177.70 TF/s | 1.76× |