Train Your TensorFlow Models on Rescale | Hacker News Reader

Train Your TensorFlow Models on Rescale

blog.rescale.com

88 pointsgpoort23 comments

Helpful FYI: If you're interested in learning about Machine Learning (so you can use TensorFlow and Rescale, etc), I've found this to be an incredible resource: https://www.coursera.org/learn/machine-learning/

If you specifically want to learn about tensorflow then you can enrol for this course by Google https://www.udacity.com/course/deep-learning--ud730

The GPU chip is Nvidia's GK104, which was released in 2012. Amazon's GPUs are similarly old. You'll train models twice as fast with a Titan X installed in your local machine, without having to deal with distributing the training.

When will we see the first cloud machine learning provider with up-to-date GPUs?

Good point. Especially for me as I'm not quite up to date with knowlege about all the GPU business (other than they're screamingly fast)

It's still scads faster than my 8 year old laptop that I'm stuck with at the moment.

Yes, I agree, that hardware is aging. We are currently working on making K80s available and should be ready in the next couple months.

(works at Rescale)

K80 is still Kepler based. Why not Maxwell?

https://thevirtualhorizon.com/2016/04/09/whats-new-in-nvidia... (tl;dr - you have to pay more to wrap the GPU access in a VM. Like, $550/more, plus a year of software support, according to that article. IANAL and haven't read any of the license details, so I don't know if there are more gotchas.)

K80s are only our next release, we do not plan to stop there.

Nervana Systems http://www.nervanasys.com/products/ is using Titan X's for their cloud.

Interesting. Does that require you to use Neon though? Also, seems like a "contact us for quote" kind of situation.

How much RAM do GK104 chips have? Wikipedia seems to imply it's about 3GB, but that's even smaller than EC2 GPU instances' 4GB. Many modern models really need the maximum 12GB to train.

Is it? Their website (http://www.rescale.com/pricing/) says they have Kepler K520 and Tesla K40. No affiliation with them at all, and haven't logged in yet, but I've been looking for something like this too, so I was hopeful until I saw your comment...

It is not the only hardware on Rescale. K520 and K40 are definitely available - after you create a free account and login there is a very transparent price list of the hardware available to you.

(works at Rescale)

Rescale pricing: "Price as low as $0.04/core/hour" - if I am reading this correctly, training on a K40 Nvidia GPU (with nearly 3000 cores) amounts to about $100/hour. Bit pricey, no? Or am I not reading this right?

the real price is "request for quote", so your number seems to be in the ballpark :)

I'd assume that is priced per CPU core (never seen anything priced per GPU core).

From memory the $0.04/hour is the base price for EC2 GPU spot instances.

Yes, you are correct. All our pricing is currently by CPU core. We are working on updates to make pricing more clear for GPU offerings.

Our rate for a single K40 at the moment is $6/hr.

(works at Rescale)

How do you plan on making your service's pricing competitive? Amazon's current spot market instances are ~30 cents per hour, which is twenty times cheaper.

For a faster GPU, that allows me to iterate more quickly, I'd be more than happy to pay an extra few bucks an hour.

I cannot comment on future per node pricing at the moment, but we are currently working on new offerings that should improve both performance (CUDA cores/node) and $/CUDA core metrics.

EC2 GPU spot instances average around $0.20 per hour in the cheapest region. On-demand instances are $0.65 per hour.

How can I install a custom python package? I've been using tflearn for my models, is there any way to use it?

Yes, you can install custom packages.

We can probably also add tflearn to the base tensorflow install so you don't have to set it up yourself.

Feel free to contact me directly at the email in my profile for more info on this.

Reply on news.ycombinator.com