I am testing cublasHgemm on 2080Ti, according to the product docs, 2080Ti has fast fp16 mode which should be 2x faster than fp32, but when I run it on 2080Ti, it did not faster. the benchmark app was compiled on 1080Ti with cuda 10.2 and then run on 2080Ti, I have add nvcc flag -arch=sm_75.
according to my test, the matrix need to be very large, then fp16 will be faster than fp32
mnicely
3
Can you provide a reproducer and results?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| why cublasHgemm is slower more than cublasSgemm when I use? | 6 | 4466 | January 22, 2019 | |
| cublasHgemm is slower than cublasSgemm in CUDA 11.1 when I use? | 2 | 574 | December 1, 2020 | |
| why is cublasHgemm is slower than cublasSgemm when matrix is low dimension | 0 | 507 | January 22, 2019 | |
| SGEMM FP16 compute? | 6 | 3974 | December 4, 2016 | |
| fp32 sgemm and fp16 hgemm | 0 | 1817 | July 4, 2016 | |
| Multiplying FP16 large matrices with cublasLtMatmul on RTX 3070 and V100 | 0 | 104 | March 31, 2025 | |
| Why does cublasSgemm uses `f16` for `float`? | 6 | 1526 | March 8, 2023 | |
| cuBLAS severe underperformance on cublasSgemm for RTX 3060 Laptop GPU | 1 | 94 | May 14, 2026 | |
| cublas sgemm,dgemm performance issue on telsa 10 and gtx 570 | 1 | 1347 | February 24, 2013 | |
| Adapt FP32 operation with TF32 | 4 | 772 | October 7, 2021 |