GeForce RTX 3050 6GB
状态与说明
Ampere die on the GA102 Booter path, not an Ada experimental target. 18 SMs. Register-resident FP16/FP32 18.07 → 36.01 TF/s, but cuBLAS GEMM gained only 22–69% depending on shape because the card spends most of a large run power-limited.
实测吞吐
| 模式 | 解锁前 | 解锁后 | 提升 |
|---|---|---|---|
f16/f16 | 36.60 TF/s | 36.60 TF/s | 1.00× |
f16/f32 | 18.30 TF/s | 36.60 TF/s | 2.00× |
bf16/f32 | 18.20 TF/s | 36.30 TF/s | 1.99× |
s8/s32 | 72.80 TOP/s | 72.60 TOP/s | 1.00× |
dp4a | 17.70 TOP/s | 17.70 TOP/s | 1.00× |
ffma | 8.40 TF/s | 8.50 TF/s | 1.01× |
cuBLAS f16/f32 4096³ | 15.83 TF/s | 22.47 TF/s | 1.42× |
cuBLAS f16/f32 8192³ | 16.05 TF/s | 19.66 TF/s | 1.22× |