I just learned about them earlier today and the service seems very interesting. Their G1 instance uses nvidia K80 (like AWS P2) and is priced at less than 50% of AWS.
I’ll be trying them out next week. If anyone already has experience with them, please share your feedback.
(I’m not associated with the company in any way)
I sent this over to Jeremy as it would be ideal for Part II.
They totally reworked the entire system, was limited to Python 3 and TensorFlow (no Jupyter Notebooks) but now it supports most every configuration. Pricing went up a bit but far more flexible. Even supports Jupyter now and will have even faster GPU’s than AWS soon.
Supports per second billing and it auto shuts off your instance after your job has run. Will cost significantly less than AWS and less cost overrun.
I still prefer to do my own hardware, but for very large jobs or if you don’t have your own GPU(s) it is a killer option.
I’d be interested in hearing feedback too!
They give you 100$ for free with second counting of time, seems like much more than we need for course!
It’s actually 100 hours free, but that’s still really generous. I’m definitely going to check it out. I’ve got a 980 in my machine which has been pretty good so far but I’m interested in seeing what you can accomplish with a k80.
The fact that they’ve set it all up specifically for deep learning is awesome, and even better they support jupyter so it’s really easy to get started.
Their GPU instances have 32 gigs of ram and 12GB K80s for $0.43/hour which seems incredibly reasonable. They also have high performance coming soon, which I’m assuming will be multi-gpu clusters.
Great find! Thanks for sharing!
The 980 will be slightly faster than the K80.
Their pricing was $.34/hr when it first came out, but they raised the price when they added Jupyter notebooks and the additional frameworks.
Hey guys - this is Sai. I’m one of the co-creators of FloydHub. Glad to see we’re being useful!
I’ll be keeping tabs on this thread. Let me know if I can answer any questions, or if you have any feedback/feature requests. You can also mail us at support@floydhub.com.
What we’re working on right now - adding support for more DL frameworks (Chainer, PyTorch, MxNet, etc.) and porting to Floyd + writing up guides for some useful DL algorithms.
Two good alternativees are also
and Microsoft Azure. Azure’s pricing is similar to AWS, but they charge in minute granularity not hourly, which can add up. Also every power cycle in AWS charges you a full hour, which is a rip off too.
Looks like you were overwhelmed by response of new users it takes really long in queue … 
Apologies! Yes, we’re experience some heavy traffic (thanks guys! ) and are overcapacity on our GPUs. Feel free to use our CPUs for free. We’re actively working on provisioning more GPUs, but till then you might see your jobs being queued - sorry! 
When I try to use jupyter notebook or upload CatsDogs dataset it shows [error] : To many open files . I guess you have some kind of blockade or serwer problem
Sai, is there anyway Floyd can partner with Fast.ai to offer free GPU access for Research and learning purposes. This is a way of giving back to the community and capacity building
Thanks for pointing that out @maciej. We’re working on improving this. In the meantime, please see our FAQ on best practices for dealing with large/many files: Too many files error, Syncing too long
@geniusgeek We are a pretty lean startup ourselves, so we don’t have the resources to offer GPUs for free 
But we understand that GPUs are insanely expensive and one of our primary goals is to lower their cost. We currently cost <50% of AWS on-demand instance, and we’ll do our best to keep it that way or lesser.
Happy to talk to folks from Fast.ai to see how we can best help!
Hey,
Thanks for sharing this resource. While I am still waiting to get AWS’s approval for my request to use P2 instance, it was natural for me to start exploring this alternative resource.
I have to admit that I am a newbie, so I had quite a number of failures to begin with. Eventually, I had it figured out - for lesson 1 and for the sample dataset. I have a rough write up of my setting up process, hoping it will be helpful and that other learners don’t have to replicate all my failures.
Here is my write up.
Any suggestions/comments on how to improve/optimize the setup process would be helpful.
cool , could you tell afterwards how it worked out? I gave up after few trials of trying setting up valid connection between data and experiment
This is an excellent writeup! Thanks for putting it together
Hi Sai,
Is the service available in India? Currently I am using aws p2 instance, but it turns out to be expensive owing to the high currency conversion rate.
Hi, good news. I got the full dogs and cats dataset figured out. Updates are reflected in the github repo:
Feel free to let me know if you have questions. I’ll try my best to figure things out.
You could use bcolz to do preprocessing locally, and then save it, and upload it.