search

ryujaehun / pytorch-gpu-benchmark

Using the famous cnn model in Pytorch, we run benchmarks on various gpu.
MIT License
226 stars 85 forks source link

My 1080 GTX Ti is 50% slower than your benchmark #3

Closed NikEyX closed 4 years ago

NikEyX commented 5 years ago

Hi,

I ran your benchmarking suite, and unfortunately it looks like your benchmarks on the 1080 GTX Ti are much better than mine. Here is a comparison for SINGLE precision on the TRAINING part:

image

Would you have any idea what causes the differences? I am basically 50% slower on all benchmarks!

I tested this on the 1080 GTX Ti and am using PyTorch 1.0.1 with Cuda 10.1 running on an AMD Ryzen 7 1700 Eight-Core Processor (with 16 threads). Running on linux.

lorenzoFabbri commented 5 years ago

@NikEyX Did you find out what the problem was? I'm having the same issue.

NikEyX commented 5 years ago

Unfortunately not. Let me know if you find a fix :)

tslaton commented 5 years ago

I'm getting similarly worse results as well with an EVGA GeForce GTX 1080 TI SC2. Temps are good, so it shouldn't be throttling or anything. Tried with a fresh conda environment with nothing installed but the dependencies and JupyterLab on both Windows 10 and Ubuntu 18.04.

Processor is an Intel i3-8350k. PyTorch version 1.2.0 and CUDA 10.0.130 (installed using default instructions on PyTorch's website).

ryujaehun commented 4 years ago

The environment I used in my experiment was: I will check again soon.

ryujaehun commented 4 years ago

First, apologize for the late response.
When I analyzed the problem, the cause was the batch size. Experimental results were run at batch size 12. However, if your results show a significant difference, it's because:

Even for this reason, it is thought that no more than 10% difference will occur.