[PYTHON] P100-PCIE-16GB was added to the GPU of Google Colab before I knew it
※※ This article is as of November 24, 2019. While thinking that everyone may already know. .. .. ..
About GPU gacha of Google colab
I think it is well known that there is a GPU gacha. See below for details. https://qiita.com/koshian2/items/d33edc963ed6cfcad77e
Trigger
I thought that it was faster than usual, and when I checked the GPU, P100 was added as follows.
GPU speed comparison
Immediately, I compared the training speed. The code used is the pix2pix code that I picked up from Google Seedbank. https://research.google.com/seedbank/seed/pixpix_with_eager_execution
Click here for Seedbank. https://qiita.com/tomo_makes/items/e5a309687f5054ba471f
Set epoch to 10 and here is the result of speed comparison.
| GPU | Calculation time/sec |
|---|---|
| K80 | 723.3 |
| T4 | 411.4 |
| P100 | 250.1 |
It's really fast. It depends more and more on Colab. .. ..
GPU gacha result (30 times)
By the way, to see how it is distributed I repeated "All runtime resets" → "nvidia-smi" about 30 times.
| GPU | Number of times | probability/% |
|---|---|---|
| K80 | 12 | 40 |
| T4 | 4 | 13.3 |
| P100 | 14 | 46.7 |
With this feeling, it seems to be about the same probability as K80.
- If you use it too often, only K80 will be assigned, so I think this result is for reference only.
Finally
The number of gachas is 30 times, but in fact I thought I would do it about 100 times.
However, on the way, I got the following allocation refusal. .. ..
I was curious about what it was, but the disastrous result was that I couldn't proceed that day anymore. Everyone, GPU gacha is moderate! Have a great Google Colab life!
Recommended Posts