You were probably downvoted because you were shilling your company, and you misunderstood the comment.
"The Osborne effect is a social phenomenon of customers canceling or deferring orders for the current, soon-to-be-obsolete product as an unexpected drawback of a company's announcing a future product prematurely. It is an example of cannibalization."
Shilling is ok on a topic directly related to my business.
You're right on the Osborne effect though! Thanks for that. We are definitely not doing that.
To clarify: When we started, MI300x was not officially announced yet, so we were planning on buying MI250's. Due to everything taking longer than expected around starting the business and receiving funding, by the time we had money in the bank, it was time to buy MI300x. Going forward, we are buying MI300x today and will continue to buy AMD MI series as they are released in the future.
Since we're on the topic of your business: I am training a decent amount of neural nets nowadays (mostly, around the new-gen robotics policies) and use vast.ai instances with 8x RTX 4090 cards.
I've been interested to give 8x MI300x a try, as they are supposed to be cheaper per FLOPs, but it looks like your service does not provide on-demand pay-per-second instances. Any plans to change that?
I would love nothing more than to be able to enable on-demand GPUs, but unfortunately this is a limitation from AMD right now. We can't do PCIe pass through to a virtual machine, it just doesn't work. This is why our minimum is 8 right now. If you look at all of our competitors, they have the same issue. Even Azure "VM", is 8 at a time, but they are all sold out due to high demand.
It kind of makes sense since their history is only supporting the high end GPUs in their HPC solutions, where they don't use VM's. They've committed to us directly that they will fix this issue.
> I would love nothing more than to be able to enable on-demand GPUs, but unfortunately this is a limitation from AMD right now. We can't do PCIe pass through to a virtual machine, it just doesn't work. This is why our minimum is 8 right now. If you look at all of our competitors, they have the same issue. Even Azure "VM", is 8 at a time, but they are all sold out due to high demand.
Thank you for the response!
Renting 8 GPUs at once is fine and desired; it's not the issue.
The issue is that right now one has to commit to at least 1 week of use; my use patterns are bursty and it does not map well to the current proposition.