Foundations Atlas

ML dev tool that saves you up to 8x in cloud GPU costs

Atlas allows you to run, track & evaluate machine learning experiments on your infrastructure.
Currently used by 700+ data scientists, Atlas helps run experiments concurrently & supports the usage of preemptible/spot instances, saving you up to 8x in GPU cost
Discussion
Would you recommend this product?
19 Reviews5.0/5
Hello friends 👋🏽, PM for Foundations here. Thank you so much @robjama for hunting us and for the support 😍😻 We started building Atlas a few years ago for ourselves when we were working with some of the largest teams to build and deploy machine learning models for enterprise. We’ve also built some crazy things on the side, which you may have seen around the interwebs (DeepFake Joe Rogan anyone?). Through our work, we’ve encountered firsthand the problems that limit ML teams impact: 1. ML teams spend a lot of time dealing with infrastructure-related work 2. Ad-hoc experiment management chews up a lot of time and limits reproducibility 3. GPU's are expensive❗️ 💰 (2.48$/hr for a NVIDIA V100??!!) 4. Teams within organizations don't usually have visibility on all data science projects and have a difficult time collaborating 5. Existing tools aren’t flexible and require dramatic changes to workflow All of this makes it difficult to be truly productive with ML. Atlas aims to solve these 5 problems. - ⚡️ Atlas sits on top of your infrastructure and has a multi-node, multi-GPU job orchestrator run & schedule jobs remotely with 1 command. - Use < 3 lines of code to record and reproduce your experiments, alongside other features like tagging jobs, recording data artifacts, and more. - 🌎 Collaborate across your organization with Atlas Projects, project-specific discussion forums, multi-tenancy, and user access controls. - 🤑 Save on GPU costs with auto-scaling of clusters and the ability to use pre-emptive/spot instances with Atlas's built-in retry mechanism. Design principles: Atlas was built to be augmentations of existing workflows as opposed to replacements, our team is obsessed with API and developer UX design, so we paid as much attention to the terminal and SDK as we did with the UI! Since the launch of our community edition 4 weeks ago, Atlas has been downloaded by over 700+ data scientists! ❤️ We are here to get your feedback so we can continue to make Atlas better. Thank you again for your support and we hope you will find value in our work. ---------- Download the free version of Foundations Atlas Join our community Slack & chat with our Engineers!
Upvote (25)Share
I'm an instructor at the University of Toronto (world leading research in AI and ML, with one of our profs winning the Alan Turing Prize in Comp Sci last year). Dessa has been partnering with the program we are running at the university to teach students how to build startups (this year in Health Care AI!). Their experience and will to teach/help has been exceptionally helpful, and their Atlas product is providing a real leg up during our product talks with the student groups. Dessa deserves tons of attention just for their willingness to help educate and dedication, let alone their amazing product. Give it a go! This is fantastic, thanks team <3
Upvote (18)Share
@jules2689 Julian! What a fun time working with all of you folks at UofT. Loved having being able to contribute to your course. Thank you very much for the kind words!!
Lightweight and flexible for an ML model tracking tool, impressive!
Upvote (11)Share
@makaed Thank you so much Oleg! Flexibility and lightweighted-nes(?) was one of our key objectives when designing Atlas.
Nice work! It's been quite the ride here at Dessa. Congrats team.
Interesting! The biggest challenge I have is compute cost (my models take hours/days to train). But spot instances are a hit and miss for me - I hate having to rerun my job all over again.. Can your "pre-emptive/spot instances with Atlas's built-in retry mechanism" help work around that?