By comparison, xAI’s supercomputer capacity isn’t public but is estimated to be around 500 exaflops.
Edit: And that’s just fucking X of all companies so imagine what an actual serious company like Google or Microsoft could have cooked up.
So while this is impressive, going CPU-only has probably resulted in a computer which costs an order of magnitude more while performing an order of magnitude less.
There is a reason the GPU architecture is still king in this aspect.
These are super computers for HPC, not “AI model training”. The link inside is saying it’s using AI to do quantum chemistry, not developing AI itself.
Which is an even bigger let-down, because that means its entire purpose is to approximate quantum computing slightly better than other classical simulators, all of which are fundamentally incapable of quantum simulation.
Background: ML can lend a higher degree of realism to QC simulation, which can be useful for experiment development due to the expense of real quantum compute time but with a lot of asterisks relating to accuracy.
Ultimately since real QC is non-negotiable for modern quantum chemistry, this super computer was likely built as a cost-saving measure that would only be justified by a lack of funding and/or affordable access to QC.
I feel threatened.
