K80 GPUs are ignored
fshriver opened this issue · 2 comments
Describe the bug
For 19.09
, tensorflow appears to ignore K80 GPUs due to them having a compute capability of 3.7. Specifically, the message returned by tensorflow is:
Ignoring visible gpu device (device: 0, name: Tesla K80, compute capability: 3.7) with Cuda compute capability 3.7. The minimum required Cuda capability is 6.0.
To Reproduce
- Pull image down.
- Start container.
- From shell run
python3
- Once inside python session, run
import tensorflow as tf
and thentf.test.gpu_device_name()
Expected behavior
Expect it to identify the K80s and attach to them, especially as the tensorflow GPU page lists hardware requirements as being only compute capability 3.5 or above (https://www.tensorflow.org/install/gpu).
Screenshots
N/A
Client (please complete the following information):
- OS: [Linux]
- Browser [N/A]
Container Version
19.09
Additional context
I'm confused on whether this is a tensorflow issue or a specific container issue, since tensorflow does not seem to indicate they've removed support for K80s.
Hello, K80 is not supported for this container.
Supported compute capability are:
- 6.0 (P100)
- 6.1 (GTX 10XX)
- 7.0 (V100)
- 7.5 (RTX 20XX)
Ah, alright. Might I suggest adding a 'Supported compute capabilities' tab under the FAQ then? Just a thought, but otherwise I'm closing this issue.