WSE 3

[MiszS]

How is this thing even real bro, a single chip that is the size of some monitors, and 1000s of times more powerful than Nvidia's best GPUs, how do you even design something like that, and why aren't more people talking about it

[AndrewAsd]

It pretty much bears no relevance or usefullness to consumer hardware. Hence most people aren't interested in it. The same way most people don't really talk about NVIDIAs dozen or so lines of data centre hardware.

[MiszS]

9 minutes ago, AndrewAsd said:

It pretty much bears no relevance or usefullness to consumer hardware. Hence most people aren't interested in it. The same way most people don't really talk about NVIDIAs dozen or so lines of data centre hardware.

Ik, but what is the point of Nvidia GPUs then, if you can get a way more power and space efficient chip to do the same task?

[Lurick]

Highly specific use cases which require customized software

vs

More commodity hardware that can be repurposed later and can run a much broader range of software.

 

Yield problems for scale

Sure, it's 57x an H100 GPU Die in terms of size. That's basically an entire wafer, if anything is bad on that wafer then what? You could cut down H100 dies and sell them for lesser cards but the WSE 3 is an entire wafer, screw up anything and the entire wafer is done and based on what I can tell they don't sell cut down versions so the entire chip is basically thrown out. I could be wrong but I can't find anything about cut down versions.

[porina]

1 hour ago, Lurick said:

Sure, it's 57x an H100 GPU Die in terms of size. That's basically an entire wafer, if anything is bad on that wafer then what? You could cut down H100 dies and sell them for lesser cards but the WSE 3 is an entire wafer, screw up anything and the entire wafer is done and based on what I can tell they don't sell cut down versions so the entire chip is basically thrown out. I could be wrong but I can't find anything about cut down versions.

Designed in redundancy. They know what error rate fabs have which isn't zero. So they build in extra sections and routing to cover for that. Find anything bad, it gets routed out and replaced. So in that sense, they're all intentionally cut down because they never intend to have a perfect wafer and use every possible part of it.

 

Link should take you to the bit covering defects. 1.5% extra cores to cover for it.