On the other hand, we've got two hyperscalers offering these things, at scale.
https://techcommunity.microsoft.com/t5/azure-high-performanc...
https://www.tomshardware.com/news/amd-scores-two-big-wins-or...
Oh and I have 3 other competitors in the smaller CSP realm offering MI300x, that I know of today.
> there are zero reliable and detailed public benchmarks on MI300x
I not very familiar with the space, but as an outsider looking in your datapoint appears to support the parent poster's opinion that the space is not mature. Wouldn't a complete lack of public benchmarks suggest that the ecosystem just isn't there yet?
My point is to agree that it isn't mature. Why isn't it mature? Well... nobody has access to these things. Today, they are only available behind HPC / hyperscaler walls. In order to gain some maturity, someone needs to offer them out in the wild. That's where I come in.
But don't take my word for it. Geohot, maybe one of the best hackers of our generation can't even it working: https://wccftech.com/tinycorp-ditches-amd-tinybox-ai-package...
It IS trash and it's clearly NOT mature. These are not mutually exclusive. There have been attempts to get AMD to open source their firmware to let people work on it since they themselves seem disinterested in doing so. Until then AMD is not a real choice for serious ML work.
Did you know... George has never used a MI300x and refused my multiple offers to let him try mine, calling me an "AMD shill." If you know me personally, that is hilariously the furthest thing from the truth one can get.
Again, before you can challenge me, I have already challenged you to back up your claim that "it is trash." Quoting George certainly is not that.