[will-burner]

Who buys chips like this and what are they used for?

It's obviously important to push the cutting edge, and it's imaginable that someday a chip like this could be in a personal computer (over 100 cores and over 4TBs of memory!).

The only things I can think, which I don't have experience with so I don't know if it's reasonable, would be high performance instances at data centers, supercomputing, and high performance main frames (is that a thing?).

[gomerspiles]

Spending this much on a server's CPUs has been a pretty normal thing in DC/server rooms for a very long time, the rise of X86 in servers is a similar cheapening like the rise of ARM. It actually meant a lot of redesigning of applications to work with many entirely separate machines instead of the smallest large server that could cover the full load and maybe around an application process per core, all sharing direct access to the same storage, etc.

[Sohcahtoa82]

> Who buys chips like this and what are they used for?

I imagine cloud service providers would buy it.

On a price/performance chart, you'd think you'd be better off buying consumer-grade Core i9 CPUs, or even i7, but that ignores everything else that must be bought, like a motherboard and RAM, and doesn't even account for the added rack space.

Put 256 GB of RAM next to that CPU and AWS can host 64 c7i.large (2 core, 4 GB) instances for $64/month each, $4,096/month. They'd see RoI in under 5 months for the CPU.

But really, this CPU is great for any embarrassingly parallel task that doesn't run well on a GPU.

[willcipriano]

When you do colocation you often pay per rack unit. Fitting more cores into a server means you need less of them and can have a lower monthly bill.

[m463]

I would use it for

  make -j 256

I'll bet folks like Linus Torvalds would get reasonable time-savings with his kernel compiles.

[rincebrain]

In any situation where power and cooling are effectively not your bottleneck, density becomes a limit on how much you can put in your server room. (Even if they eventually become your bottleneck, you might conceive of a situation where you build out to 75% capacity in 50% or less of the space to let you put in more efficient future versions in the other physical volume, or something.)

I would imagine any of the big cloud providers, for the ones providing x86, would probably love to buy something like this to improve the density of what they can offer...

[chaosbutters314]

i did. small scale (cfd/fea) simulation when the cloud or company server doesnt make sense. with how much data the model makes, easier to work locally and test it runs before spending time and money on cloud. also, a lot of simulatione dont need to be that big so this is great for working all locally and not having to pay core hours. usually pays for itself in a year frankly