mila-iqia / milabench

Repository of machine learning benchmarks
https://milabench.readthedocs.io
MIT License
19 stars 23 forks source link

Jax & CUDA #258

Open Delaunay opened 3 weeks ago

Delaunay commented 3 weeks ago
=================
Benchmark results
=================
bench                          | fail |   n |       perf |   sem% |   std% | peak_memory |      score | weight
brax                           |    1 |   1 |    4822.59 |   0.6% |   1.4% |         nan |       0.00 |   1.00
diffusion-gpus                 |    0 |   1 |     360.89 |   0.1% |   1.0% |       60298 |     360.89 |   1.00
diffusion-single               |    0 |   1 |     362.63 |   0.1% |   1.0% |       60336 |     362.63 |   0.00
dinov2-giant-gpus              |    0 |   1 |     864.20 |   0.4% |   2.9% |       72102 |     864.20 |   1.00
dinov2-giant-single            |    0 |   8 |     102.86 |   0.4% |   9.5% |       74544 |     831.34 |   0.00
bf16                           |    0 |   8 |     784.40 |   0.2% |   5.8% |        1548 |    6308.94 |   0.00
fp16                           |    0 |   8 |     799.06 |   0.2% |   6.3% |        1548 |    6431.99 |   0.00
fp32                           |    0 |   8 |      51.94 |   0.0% |   0.7% |        1932 |     415.81 |   0.00
tf32                           |    0 |   8 |     406.56 |   0.2% |   5.5% |        1932 |    3258.38 |   0.00
dimenet                        |    0 |   8 |     483.93 |   0.7% |  15.0% |        7456 |    3905.80 |   0.00
bert-fp16                      |    0 |   8 |     457.20 |   0.9% |  14.2% |         nan |    3744.58 |   0.00
bert-fp32                      |    0 |   8 |     111.46 |   0.4% |   6.5% |        6436 |     901.45 |   0.00
bert-tf32                      |    0 |   8 |     243.67 |   0.7% |  10.3% |         nan |    1983.29 |   0.00
bert-tf32-fp16                 |    0 |   8 |     457.23 |   0.9% |  14.3% |         nan |    3744.93 |   3.00
reformer                       |    0 |   8 |     103.20 |   0.4% |   8.2% |       25208 |     833.78 |   1.00
t5                             |    0 |   8 |      89.13 |   0.4% |   9.7% |        9208 |     721.16 |   2.00
whisper                        |    0 |   8 |     878.49 |   0.5% |  10.7% |        1026 |    7113.36 |   1.00
lightning                      |    0 |   8 |    1222.54 |   0.6% |  13.7% |       27078 |    9900.16 |   0.00
lightning-gpus                 |    0 |   1 |    9849.93 |   0.7% |   5.7% |       30516 |    9849.93 |   1.00
llama                          |    0 |   8 |     768.96 |   4.8% |  85.9% |       27608 |    5787.01 |   1.00
llm-full-mp-gpus               |    0 |   1 |     513.47 |   2.4% |  12.9% |       49382 |     513.47 |   1.00
llm-lora-ddp-gpus              |    0 |   1 |   29652.09 |   0.4% |   2.2% |       39352 |   29652.09 |   1.00
llm-lora-mp-gpus               |    0 |   1 |    3810.83 |   1.8% |   9.3% |       55454 |    3810.83 |   1.00
llm-lora-single                |    0 |   8 |    5048.74 |   0.3% |   4.2% |       49732 |   40488.86 |   1.00
recursiongfn                   |    0 |   8 |   12029.94 |   0.8% |  17.8% |       10692 |   97098.93 |   0.00
super-slomo                    |    0 |   8 |      87.64 |   0.7% |  14.5% |       67926 |     706.49 |   1.00
focalnet                       |    0 |   8 |     649.43 |   0.6% |  13.0% |       24524 |    5263.90 |   2.00
torchatari                     |    0 |   8 |    8370.17 |   0.3% |   7.1% |        3690 |   66911.20 |   0.00
convnext_large-fp16            |    0 |   8 |     663.05 |   1.0% |  15.2% |         nan |    5440.84 |   0.00
convnext_large-fp32            |    0 |   8 |     129.51 |   0.6% |   9.4% |       47170 |    1052.58 |   0.00
convnext_large-tf32            |    0 |   8 |     239.24 |   0.8% |  12.6% |       49506 |    1955.14 |   0.00
convnext_large-tf32-fp16       |    0 |   8 |     664.53 |   1.0% |  15.2% |         nan |    5452.64 |   3.00
regnet_y_128gf                 |    0 |   8 |     187.71 |   0.4% |   9.0% |       28740 |    1517.87 |   2.00
resnet152-ddp-gpus             |    0 |   1 |    8172.97 |   0.1% |   0.4% |       30964 |    8172.97 |   0.00
resnet50                       |    0 |   8 |    1804.63 |   0.6% |  13.0% |       13340 |   14621.76 |   1.00
resnet50-noio                  |    0 |   8 |    2029.43 |   0.4% |   7.8% |       27354 |   16386.23 |   0.00