New Tesla T4 available in google collaboratory!

fastai
21

Thanks for the answer. And no problem at all. Copying large archives (tar balls) across from a gcs bucket to to colab with gsutil cp ... is indeed fast. Not sure about latency with copying lots of smaller files, though.
The thing that has not worked for me is mounting the bucket to a mount point (so like a drive). I get reading directory '...': Input/output error errors when trying to access the mounted bucket. That’s why I asked if you were able to mount the bucket as a drive.

22

There is a cuda dataframe library that runs data loading and processing on the GPU.
However, it requires a Tesla GPU with tensor cores and I believe the new T4 GPU on Colab fits the bill.

If your data-loading is CPU bound and slow, that lib might does the trick.

Benchmarks:
https://matthewrocklin.com/blog/work/2018/12/17/gpu-python-challenges

1 Like
23

I think is dual-core not single.

24

Colab and Kaggle Kernels are both dual cores, but in Kaggle Kernels there is an issue, at least in Pytorch, that you cannot use multipple ( in this case 2 ) threads Dataloader. Because of this, in my experience, Kaggle is 2/3 slower than Colab.

25

nvm I had my num_workers=0. Increasing to 4 leads to a 2x increase in speed compared to kaggle kernels! Yay!

1 Like
26

Works only with new accounts. On my old account I always get a K80, but with a new one I got a T4, too. Yay :).

27

I have tried it with my existing account and it showed me T4 and TPU options. I didn’t have to create a brand new account to see the T4 option. Maybe they’ve only rolled it out to only a segment of old accounts but all new accounts get a T4/TPU option.

28

I have a feeling they maybe sharing a T4 amongst 2-4 colab kernels.

29

Yeap, it is probably the case right now. It is painful slow. The kaggle P100 looks way more interesting now.

30

Hi guys! I have a 1060 6gb, so is it faster with collab T4 or 1060 6gb?

31

probably your dedicated 1060 will be faster. You will have to try.
Anyway, you have $300 credits free with GCP and you can test a proper dedicated T4.

32

+1 on Phucnsp’s comment, I am unable to access the forum you started, any change you can make this public?

33

@mike.moloch try clouderizer.com with your old account. It will give you T4 by default.

1 Like
34

Does anyone else still get the K80 when they use the GPU runtime for Colab? I can’t seem to find a resource on how to switch to the T4s

1 Like
35

Sometimes yes. I guess it’s an availability issue given Colab is getting more and more popular.

36

you could try nvidia’s dali library, which also allows to run jpeg decoding & image augmentation on gpu.

37

me too…