I started my deep learning journey 6 years back. Fastai along with its amazing community played a crucial role in building the skills that helped me in winning various Kaggle competition medals. It has become a practice for me, that I wait every year for the next fastai course and I learn a lot of new things each year. Last year I quit my job and started exploring new opportunities.
Becoming an entrepreneur was a long time dream. During all these years of studying and practicing deep learning, I came across multiple GPU cloud platforms. Each of them offered different features like simplicity, increased affordability, wrappers around cloud platforms. I was an early adopter to some of the platforms like Salamander, Vast.ai, and Crestle.
Inspired by these platforms, I aspired to build a cloud platform that is simplest to use and most affordable for deep learning users. I am very excited and happy to share with you all that our platform is available for use at cloud.jarvislabs.ai and we are happy to offer an initial 20$ credit (close to 50+ hours of GPU compute) to use the platform. Find the demo video here
Currently, the platform supports features like:
Supports popular frameworks like Pytorch, Fastai, and Tensorflow
Multi GPU option.
Currently we are offering RTX 5000 & RTX 6000 graphic cards.
One-click Jupyter notebook and ssh access (If the public key is provided)
GPU price starts as low as 0.39$.
Some of the features that we will be adding in the future are.
Ability to pause and resume.
Alert when the machine is idle for a long period.
Will be adding A6000 cards.
As an early-stage startup, we request you all to try our platform and share your feedback on usability, prices, and features. If you like the platform please spread the word among your friends and colleagues.
Thank you,
Vishnu Subramanian
**Note We have stopped providing credits on SignUp since a lot of folks started using disposable and fake ids to take advantage of the free credits. We have decided to provide credits only after verifying the ids so that our genuine customers are not affected. So if you are signing up, please drop an email to hello@jarvislabs.ai with the email id used for signup and your fastai profile link. You can also ping us on the chat option that you find at the bottom of the screen.
Hi, I like this idea and will give it a try.
I have 1 question only: can u give selectable HDD size too?
I mean I saw there is fixed 100 GB for 1 gpu, 200 GB for 2 gpus, … 400 GB for 4 gpus, but then there is the max 400 GB HDD till 8 gpus.
Some kaggle competion has 2-20 GB dataset, but I had 1 competion that has 1 TB, the “RSNA STR Pulmonary Embolism Detection”.
That’s why I ask for selectable option there - 90% of the time hundreds GB enough, but sometimes not.
I get the point. I have faced the same problem with some of the object detection competition hosted on Kaggle. We have some features that we want to implement where we can select the amount of hard-disk and that persists across machines.
Okey, I signed back in and now writing to you there.
I tought as a 2nd guess maybe the problem was my browser, you know maybe cookies or java script etc - I use Vivaldi, but then that was not the problem.
Previous time it was dawn and it was my 1. try to sign up, etc.
I didn’t tried the training yet, but I write here my first expressions about the project/website.
So the 1. good thing there is fast supportive response and problem solving
Some browser like my Vivaldi automatically block trackers and ads, so on this site it blocked the support chat button in the right bottom corner and it was not visible at all.
I realized maybe elements on the site blocked and unblock it - then I saw the chat button.
My suggestion to put 1 normal html button for FAQ to the site and write there:
“If you do not see support chat button in the right bottom corner - then please check your browser maybe blocks it, and you need to unblock it”.
It can help other people if step in my case ^^
The site is pure minimalistic, really just for 1 single goal: start your training instance fast and that’s it
I started an instance, it was indeed fast - actually faster than the promised max 30 seconds in the tutorial video - it was 5 seconds for me.
And the site shows the running instance time in hours and minutes - I just saying this, because I like exact times and for example Paperspace shows only integer hours without any minutes.
I asked Paperspace maybe there is a terminal command to query exact running time - but they said there is no such feature on their platform, which is strange for me. Here it was default
The price and gpu power seems competitive here and I got the promise for faster internet connection at end of week, so this can be a good alternative to train models.
I’ve tried the platform a couple of weeks ago and the signup process and the interface was really nice and smooth, without unnecessary cruft. You pay and you can spin up Jupyter notebooks with one click ASAP. You even get SSH access out of the box (if you add your public SSH key in the settings). The prices are great and the machines seem be equipped exactly as advertised.
Two things that are problematic:
The machines can be stopped but they disappear (together with all the code and data) after a variable number of hours (48-72 AFAIR). It would be nice if there was at least an e-mail reminder on the day this is going to happen. It would be even better if you could pay a small fee to keep files alive between sessions.
The machines seem to have no shared storage and you cannot easily change the GPU allocation on an existing machine (even when it’s stopped) so you may be stuck with the selection you started with.
The download performance was very bad. I spent 3 hours downloading a 20GB Kaggle dataset. I did not want to leave it overnight (to avoid paying for a unused machine) but then I could not resume the download (Kaggle’s fault, not Yarvis’). Fortunately I was able to download it quickly to my computer and fix the partially downloaded file using rsync. But the fact that I was able to saturate the bandwidth using my cable modem at home means the network performance is pretty bad.
Thanks for a lot of great work @VishnuSubramanian and I hope my review helps you guys decide if it fits your use cases.
Thank you for using our product and providing valuable feedback. We are working hard to make the platform more robust. We are working on some of the features which will address the challenges you have mentioned in the coming weeks. We will post an update here soon.
.
We have increased the paused instance duration from 2/3 days to 7 days (168 hours). So all the data, and code would remain intact for 7 days. After that, we will delete the instance.
We are exploring options where we could allow users to pay only for the storage and not worry about termination. So all the data and code could remain as long as you want(Till the account has sufficient funds). We should be able to implement it in the coming weeks.
The ability to change the type of GPU and the number of GPUs while resuming paused instances is also part of the product roadmap. You should be able to do it soon. It is a feature that I would love to have as a deep learning practitioner.
I apologize for the download performance. We were promised by our Internet Service providers for faster internet by December 2020. We are hoping we will get a faster internet speed in the next few days. We will update you once the internet speed is upgraded.
We are a small team and working on improving the product every day. Feedback like this means a lot to us. Thanks for sharing and we would love you to spend more time training models on JarvisCloud.
