Top Cloud GPU Providers For Machine Learning in 2022
Ten great providers, over $900 in free credits
If you’ve ever done any serious Machine Learning and don’t have access to a big GPU cluster, you’ve probably used some cloud GPU services.
Cloud GPUs are ideal if you’re looking for intense use in short bursts — for example, if you want to quickly fit hyperparameters across a large space.
Or maybe you’ve got a promising proof-of-concept model for a Machine Learning Competition, and want to train it on the full data set.
Whatever you need it for, Cloud GPUs can be expensive, and there are lots of options. In this article we’ll cover 10 of the top services — for a more detailed feature grid and pricing for specific GPU instances, check out our open source Cloud GPU Comparison page.
Genesis Cloud
Genesis Cloud started as a crypto mining business, and pivoted to offering cloud GPUs for ML. Their data center is based in Iceland and exclusively uses hydroelectric and geothermal energy. They have a limited selection of GPU options, but very competitive pricing.
- Regions: Iceland
- GPU models: GeForce RTX 3090, RTX 3080, GTX 1080Ti
- One-click notebooks: yes
- Persistence: yes
- Pricing: competitive 💲
- Free credits: $50
DataCrunch
DataCrunch are another specialised cloud GPU provider, this time based in Finland. They have competitive pricing, and a useful comparison page, where you can compare their offering vs GCP/AWS/Paperspace/LeaderGPU.
- Regions: Finland
- GPU models: A100, RTX A6000, V100
- One-click notebooks: yes
- Persistence: yes
- Pricing: competitive 💲
- No free credits
Paperspace
Paperspace offer virtual servers with GPUs as part of their ‘CORE’ offering. Alongside this they also have a ‘Gradient’ offering, for a more managed/high-level ML workflow. Their pricing tends to be towards the higher end, but their offering can be easier to use than others if you’re not a fan of the command line.
- Regions: US/Europe
- GPU models: Quadro M4000, P4000, P5000, P6000, RTX4000, RTX5000, Ampre A4000, A5000, A6000, A100.
- One-click notebooks: yes
- Persistence: yes
- Pricing: mediocre 💲💲
- Free credits: $10
Vast.ai
Vast.ai are a bit special — they are a peer-to-peer GPU network, which allows anyone to rent out their GPU to others who need one. This means their pricing is often cheaper than others, but since you get a different GPU every time it means you can’t store data there when you’re not using the service.
- Regions: worldwide 🌐
- GPU models: many!
- One-click notebooks: yes
- Persistence: no
- Pricing: competitive 💲
- Free credits: a few minutes of use
Jarvis Labs
Jarvis Labs are a specialised cloud GPU provider based in India. They have competitive pricing across a few GPU models, but don’t offer any free credits.
- Regions: India
- GPU models: RTX5000, RTX6000, A6000, A100
- One-click notebooks: yes
- Persistence: yes
- Pricing: competitive 💲
- Free credits: none
LeaderGPU
LeaderGPU are a specialised cloud GPU provider based in the Netherlands, and offer a wide range of GPU models at fairly competitive prices. They seem to be targeted more towards enterprise customers than individual researchers, and don’t offer a simple one-click notebook solution.
- Regions: Netherlands
- GPU models: many!
- One-click notebooks: no
- Persistence: yes
- Pricing: competitive 💲
- Free credits: none
Lambda Cloud
Lambda are an ML hardware company based in the US, and sell servers as well as having their own GPU Cloud offering. They are also the creators of the Lambda Stack package, which makes it much easier to install deep learning frameworks on any VM.
- Regions: US
- GPU models: RTX A6000, Quadro RTX 6000, V100
- One-click notebooks: yes
- Persistence: no
- Pricing: competitive💲
- Free credits: none
AWS/GCP/Azure
I’ve bundled these three together since you are likely familiar with them already. For those who don’t know — AWS is Amazon Web Services, GCP is Google Cloud Platform, and Azure is Microsoft’s cloud offering. All three are fairly mature, and GPU instances are just one of many things they offer. These might be a good fit if you already have other services running with them. They tend to be more expensive than the others on this list, but they also offer more free credits when you start.
- Regions: worldwide 🌐
- GPU models: many
- One-click notebooks: yes
- Persistence: yes
- Pricing: expensive 💲💲💲
- Free credits: $200 (Azure), $300 (AWS), $350 (GCP)
Other resources
Some other useful resources:
- To choose the right GPU, you need to take into account what type of model you want to train. Tim Dettmers has a useful in-depth analysis of the best GPUs to use for deep learning, where he talks about this in detail.
- Jeff Hale has an interesting post where he benchmarks training time for multiple different models across a bunch of Cloud GPU instances.
- Google Colab can be useful if you want to get free GPU access for prototyping!
- Our Cloud GPU Comparison page, which allows you to look at pricing for specific GPU instances across all the providers mentioned in this article.
Anything we’ve missed? Let us know in a comment!
