On traditional cloud platforms, each developer uses a server with its own dedicated GPU. On Thunder Compute, developers share a pool of GPUs, which are network-attached to CPU-only cloud instances on demand. In this system you do not pay when GPUs are idle.
My friend Brian and I were frustrated with the experience of developing with GPUs and started building Thunder Compute as a side project during college. Brian's research lab used Excel to sign up for GPUs two weeks in advance which got us thinking about better ways to allocate GPUs. We began building the GPU virtualization that became the backbone for Thunder Compute.
Two years later we went through Y Combinator, are launching, and are excited for y'all to try what we built! Thunder Compute was really tough to build and there's a lot going on behind the scenes to make this system work. We appreciate your patience with any bugs and your feedback is incredibly helpful.
Here are some of the benefits of GPU virtualization:
1. We share GPUs across many cloud instances to lower cost
On GCP/AWS/anywhere else, if you have 100 GPU cloud instances you have 100 GPUs. On Thunder Compute, if you have 100 GPU cloud instances you share ~10 GPUs. This reduces cost by improving GPU utilization.
2. You only pay when GPUs are active
Following this first point, because GPUs are shared, we only charge you for time that you actually use them. On other platforms, you pay for the GPU even when it is idle, such as time you spend developing, debugging, or accidentally leaving instances on. On Thunder Compute we can share the GPU during this time and you don't have to pay.
3. You can easily switch GPUs and don't have to install NVIDIA drivers
Another nice benefit of network-attaching GPUs is that if you want to change the GPU that is attached to your instance you simply run a command and say you want the new GPU. You don't have to move any of your files, provision a new instance, or change any settings. Also, because there is no physical GPU on your machine, you don't need to install GPU drivers, saving even more time.
Try it today at thundercompute.com, Brian and I love to hear your feedback and appreciate every one of you!
This would've saved us so much on those ML experiments last month - watching idle GPUs burn through budget hurts. Love the network-attached concept... any latency impact though? Working on some real-time inference stuff and milliseconds matter.
Congrats on the launch! This sounds like a game-changer for AI/ML devs. How do you see it comparing to existing cloud GPU options in terms of cost and performance?
It looks like a magic! 🪄 This is exactly what I need and what I will definitely try. Getting access to the NVIDIA A100 for a short time is always a pain for me - it's time consuming to get it and always expensive. The main issue for me in such a service - spendings control (limitation of planned expenses). Do you have that feature already? If not, are you planning to make it? Also it's very interesting to learn about your business model in general. Good luck with your project! 🤞
@carl_petey congrats on your launch! Very strong offering - just curious, how do you finance free GPUs or what is your business model?
About Thunder Compute on Product Hunt
“Network-attached GPUs that disconnect when idle”
Thunder Compute launched on Product Hunt on November 18th, 2024 and earned 62 upvotes and 11 comments, placing #14 on the daily leaderboard. On traditional cloud platforms, each developer uses a server with its own dedicated GPU. On Thunder Compute, developers share a pool of GPUs, which are network-attached to CPU-only cloud instances on demand. In this system you do not pay when GPUs are idle.
Thunder Compute was featured in Software Engineering (42.4k followers), Developer Tools (511.2k followers) and Artificial Intelligence (466.6k followers) on Product Hunt. Together, these topics include over 160.2k products, making this a competitive space to launch in.
Want to see how Thunder Compute stacked up against nearby launches in real time? Check out the live launch dashboard for upvote speed charts, proximity comparisons, and more analytics.
