[Solved] Strange Results using Tensorflow with GTX 1080

Never mind, its resolved

Were you able to get CUDA 8.0 installed or are you still using 7.5 for this? Just checking.

Using 7.5, the issue wasn’t actually resolved. It stopped detecting the GPU and the code got correct results when running on the CPU.

After rebooting, Tensorflow detects the GPU but has the same issues as before.

I’m also on CUDA 8 and it works.

Can you guys elaborate on your experience with the 1080? Performance?

It seems there isn’t really any data on the net about how these perform with TensorFlow or any of the ML/DL stuff yet. Would love to hear your feedback!

Initial results are pretty promising:

This was included at the top of the alexnet_benchark in tensorflow:

“”"Timing benchmark for AlexNet inference.

Run on Tesla K40c: 145 +/- 1.5 ms / batch
Run on Titan X: 70 +/- 0.1 ms / batch

Forward-backward pass:
Run on Tesla K40c: 480 +/- 48 ms / batch
Run on Titan X: 244 +/- 30 ms / batch

With the 1080, I get:
2016-06-04 21:08:10.885878: Forward across 100 steps, 0.024 +/- 0.003 sec / batch
2016-06-04 21:08:19.105703: Forward-backward across 100 steps, 0.070 +/- 0.007 sec / batch

The benchmarks in the comments were probably run with CUDA 7.5, so some of the performance gain may be due to the upgrade from CUDA 7.5 to 8 (and CuDNN from V4 to V5)

Regardless, it’s pretty impressive that we are beating the Titan X by that much.

You could have at least converted the units and labeled the data such that comparisons can be made ;)

In any case this is apples and oranges comparison because you’re changing both software and hardware.

I just reported the output and a comment that came with tensorflow. It shouldn’t be that hard to convert from seconds to milliseconds.

According to:

CUDA 8.0 is needed.