It’s a nice easy unit to compare against because all of the ones I’ve seen draw basically exactly 1000 watts
It’s also less than double what my desktop draws (11700k, rtx 3060) and (those aren’t particularly demanding components, and I only get 16 threads) (that’s basically exactly 6x more power draw per core, although the cores themselves perform differently ofc)
It is slightly silly to have that many cores tho. I guess the main reason to not just use a gpu would be because pcie doesn’t have enough bandwidth, or if you need a ton of RAM? For a pure compute application I don’t think there are many cases where a GPU isn’t the obvious choice when you’re going to have almost 400 threads anyways. An A100 has half the tdp and there’s no way the epyc chip can even come close in performance (even if you assume the cpu can use avx512 while the gpu can’t use it’s tensor cores, it having about a third of the memory bandwidth isn’t exactly encouraging about the level of peak compute they’re expecting)
I stopped reading your comment when you suggested calling 1000w as a toaster. Most murican thing. Why don’t you also measure the size of the CPU in buses?
Only half of a toaster (in the US at least)
It’s a nice easy unit to compare against because all of the ones I’ve seen draw basically exactly 1000 watts
It’s also less than double what my desktop draws (11700k, rtx 3060) and (those aren’t particularly demanding components, and I only get 16 threads) (that’s basically exactly 6x more power draw per core, although the cores themselves perform differently ofc)
It is slightly silly to have that many cores tho. I guess the main reason to not just use a gpu would be because pcie doesn’t have enough bandwidth, or if you need a ton of RAM? For a pure compute application I don’t think there are many cases where a GPU isn’t the obvious choice when you’re going to have almost 400 threads anyways. An A100 has half the tdp and there’s no way the epyc chip can even come close in performance (even if you assume the cpu can use avx512 while the gpu can’t use it’s tensor cores, it having about a third of the memory bandwidth isn’t exactly encouraging about the level of peak compute they’re expecting)
Maybe its meant for hyperscalars who will rent it out in smaller units of say 16, 32, and 64 core instances to customers.
Another application of those things are virtualization. Throw in 3 x4 TB NVMe SSDs, 384 GB of memory, and a 25G NIC.
Off a single unit, you would be able to sell 12 VPS instances with 16 cores, 32 GB of memory, 1 TB of storage, and a guaranteed 1.5Gbps link.
I stopped reading your comment when you suggested calling 1000w as a toaster. Most murican thing. Why don’t you also measure the size of the CPU in buses?
Intuition