Hacker News Reader

Comment by mikeweiss | Hacker News Reader

When are we likely to see MI300 available for rent and what would the cost per hour be?

Thanks for asking.

Soon! Estimated ship date is 2/29. Should arrive around 5 days after that. We are going to fly to the data center, get the boxes all set up and start to launch things.

Pricing is TBD since there isn't really a comparable and we expect that since these GPUs are so new, people will want to do a lot of testing before they are willing to commit. Therefore, we are looking for partnerships over just customers.

Disclosure: Check out my profile, I'm building a white glove bare metal service that is focused on only top end AMD compute and specifically the MI300x (and future generations). Feel free to reach out via email.

We can't quite afford an 8x bare metal instance, but I am pinning this for later.

Think ya'll will ever offer single MI300X instances?

Yes! Like I said above, we fully expect people to be toe dipping at first given how new this hardware is.

We offer single GPU access through a VM and PCIe pass through. Either we can provide you a basic Ubuntu VM with everything loaded into it (with regards to drivers and such) or you can send us your VM image and set it up however you like. Then you'll just have direct ssh access into it.

OK that sounds lovely. We are but a poor startup, but I am hoping we can afford to reach out to y'all soon. Maybe even share workflows/stuff that works for us. If some ML (unsloth? lorax? sglang?) work on a Mi300 with a little tinkering, that would be spectacular.

Thanks! Again, partnerships over customers. If you're experienced and have the technical chops to make a MI300x sing, we want to work with you. Our model is that we are the capex/opex investor for businesses. As much as I love software, Hot Aisle is more of a hardware business. Running super high end large scale compute is an extreme challenge in itself. We are less interested in building the software side of things and want to foster those who can focus on that side.

https://github.com/unslothai/unsloth/issues/160

https://github.com/search?q=repo%3Apredibase%2Florax+rocm&ty...

https://github.com/sgl-project/sglang/issues/157

https://github.com/casper-hansen/AutoAWQ (supports rocm)

Which companies that already rent out H100s would risk the ire of nvidia providing them less capacity next time?

edit: It seems cudocompute and dataknox is willing to rent them out

MI300 is early access in Azure right now -- https://techcommunity.microsoft.com/t5/azure-high-performanc...

This sadly doesn't mean that they are actually available. I haven't seen anyone saying they are using them. Obviously, that doesn't mean much either, but I'd at least hope to see someone post some performance benchmarks.

Early access means NDA.

Well, that's no fun at all.

Reply on news.ycombinator.com