GeForce RTX 4070 (AD103)
状态与说明
Then-untested AD103 die. The register-level probe went from half to full rate and doubled FP16/BF16 with FP32 accumulate; cuBLAS GEMM gained less on small shapes (+43% at 1024, rising to +50% at 4096) because those runs are not tensor-bound.
实测吞吐
| 模式 | 解锁前 | 解锁后 | 提升 |
|---|---|---|---|
f16/f16 | 132.60 TF/s | 132.20 TF/s | 1.00× |
f16/f32 | 66.30 TF/s | 132.20 TF/s | 1.99× |
bf16/f32 | 66.30 TF/s | 132.00 TF/s | 1.99× |
s8/s32 | 265.10 TOP/s | 263.70 TOP/s | 0.99× |
dp4a | 65.80 TOP/s | 65.70 TOP/s | 1.00× |
ffma | 31.70 TF/s | 31.70 TF/s | 1.00× |
tf32 | 33.10 TF/s | 65.80 TF/s | 1.99× |
u4/s32 | 527.20 TOP/s | 525.70 TOP/s | 1.00× |
cuBLAS f16/f32 1024³ | 49.06 TF/s | 70.14 TF/s | 1.43× |
cuBLAS f16/f32 4096³ | 64.04 TF/s | 96.33 TF/s | 1.50× |
cuBLAS f16/f32 8192³ | 64.42 TF/s | 93.63 TF/s | 1.45× |