[Beowulf] Register article on Epyc

Bill Broadley bill at cse.ucdavis.edu
Thu Jun 22 20:44:39 PDT 2017

On 06/22/2017 08:21 PM, Kilian Cavalotti wrote:
> Oh, and at least the higher core-count SKUs like the 32-core 7251 are
> actually 4 8-core dies linked together with a new "Infinity Fabric"
> interconnect, not a single 32-core die. I completely missed that. And
> it's fine, it probably makes sense from a yield perspective, but

Indeed, actually it's a single piece of silicon for the Ryzen (one die),
Threadripper (2 die), and Epyc (4 die).  That way AMD can amortize the R&D, not
to mention testing, compiler tweaking, etc over a larger number of chips.

The IF (Infinitiy Fabric) allow each lane to be PCI-E or a coherent memory link,
not sure if they will open it up to accelerator makers.  I was pretty surprised
to hear that when HT was a big advantage to AMD that they didn't allow licenses
for the coherent part.  I believe the Zen core IF interconnect on chip is fully
connected (all pairs of CPUs have a link), and then for a dual socket there's an
cross socket link from each die of socket 1 to each die of socket 2 (4 links
total).  So socket 1 chip A has a link to socket 2 chip A.

> behold the intra-socket inter-die NUMA effects. And sure enough, we'll
> need yet another layer in the node-board-socket-core-thread conceptual
> cake.

Dunno, multiple chips are pretty common.  The previous generation Opterons had
it and it was common way back to the pentium pro if not before.  Even with out
multiple chips per package there are NUMA differences within a single chip.  I
believe Intel's smaller chips are fairly simple with a double ring bus up to 8
cores, but above that they add another pair of ring busses for the additional cores.

So sure, yesterdays quad socket is somewhat like today's dual socket, but quad
socket was hardly the largest linux box you could buy.  So I think the
node-board-socket-core-thread cake is fine.

> By the way, isn't actually what Intel was criticized for, when they
> released their first quad-core Clovertown CPUs, back in 2007? I
> remember some PR from AMD about those fake quad-core CPUs, which
> really were just two dual-core glued together. Funny how tides turn...

Indeed.  I hope it works out well for AMD, being able to target 3 different
markets (desktop, HEDT/workstation, and servers) with the same chip seems like a
good way for the AMD to compete with Intel's advantage in R&D and fabs.

More information about the Beowulf mailing list