I was looking for a decent AMD replacement for a Intel NUC and didn't found it then, wanted to suggest a new niche in the previous comment ( looked for it a month or so ago). I only found out after my comment it was coming and didn't want to remove my comment ;). As i'm truelly interested in a AMD replacement for my current slow ( and cheap ) Intel NUC
:)
And euh, looking at the top comments. I'm not alone, sarcasms aside
Ontopic:
If you order enough of them. They can address a custom team for you! You could earn money like hell launching this on the market!
https://www.amd.com/en/products/semi-custom-solutions
Build your dream!
You're right to express some concern -- practical residential limit for continuous load on a perfect incarnation of the typical residential circuit is something like 1440W total (80% of nominal 15A breaker on 120VAC).
Maybe more realistic mid-range products are 155-180W/socket, and that drops the CPU draw down to 620-720W.
Yep! The typical US residential circuit is only 120v, 15A breaker (so 12A at 80% load). (And that may be only 110-120v.) 20A circuits are also fairly common, but it's not the most common; the majority of wall outlets in a US house will be 15A.
In this chart, https://en.wikipedia.org/wiki/NEMA_connector#/media/File:NEM... , the commonly used US residential sockets are NEMA 5-15 (labelled "Typical Outlet," for 15A breaker circuits), NEMA 5-20 (for, duh, 20A breaker circuits), and NEMA 1-15 ("Old Outlet," in older buildings).
> 1500W isn't that much of a problem. Oven breakers are even 20 or 32A.
We do have higher amperage circuits for appliances like ovens and clothes driers, but they're usually dedicated circuits and have different socket shapes. You can see the labels "Clothes Dryer" and "Electric Oven" in the chart linked above :-).
And of course some datacenters just wire DC to the whole schebang instead of having an individual, well two redundant, PSUs per server.
AFAIK, there aren't any, Epyc only has 1 socket (with 129 PCIe lanes) or 2 socket (using 64 or 48 of the 129 lanes to connect to the other socket) options.
However, 'chrome' would be valid. :)
There's algorithms that haven't yet been optimized for GPU parallelization. And there's datasets that require memory access patterns that aren't amenable to GPU.
That said, now I'm curious what a 128-core system would do to memory access. Better hope the algorithms are L2/L3 coherent?