Intel Xeon 6980P vs. AMD EPYC Power Efficiency / Performance-Per-Watt Benchmarks

[ET3D]

Nice showing from Intel. Sure, AMD's next gen is coming up, but it's still nice to see Intel back in the ballgame and on it's own process.

are we supposed to believe that Intel 3 is competing with TSMC 3nm

Far as I remember Intel expects 18A to be on par with TSMC 2nm and TSMC expects it to be on par with the latest 3nm.

It's worth remembering that two process characteristics are density and power use, and that "3nm" by itself is meaningless. Even at TSMC, N3E is more power-efficient than N3 (N3B), and N3P is even more power efficient and also a little denser. It's possible that Intel 3 is as dense as N3, but not as power-efficient, or that it is, but not as efficient as N3E.

Short of it, it looks like Intel is back in the game, but it's hard to tell what exactly to compare it to.

[geerge]

The node names are mostly marketing bullshit these days anyway, but you are putting to much weight in Intel's "3" being competitive with TSMC's "3". Intel renamed their old node targets to better align with TSMCs. Intel 4 = their old 7nm node. Intel 3 is an enhanced version of that. It isn't in any way efficiency competitive with TSMC's leading edge node.

I'm not putting any stock in the name "3", what I'm comparing is what intel is offering as their current (I know they're iterating quickly) leading node with the competition. It's not efficiency competitive as you say and that's a red flag. Supposedly 18A is much better and "3" is a short-lived interim step, but it's the best they're currently offering so it's fair to compare and find out where they're at.

Yeah, but you are then comparing AMD's E-cores( Zen 4c) to Intel's P-cores. Neither is a great direct comparison of 1:1. Depending on the test all 3 had good/impressive showings though.

I largely disagree when it comes to server. When it comes to desktop yes maybe comparing big to little is less relevant, the gulf in power consumption between the two is large and there are other factors. So intel has a smaller core and AMD has a bigger core not being compared, big deal. If anything the 9754 vs 6980P comparison is more valid because intel has double the cache which eats into the power budget, bringing the consumption of the cores on both parts closer together than it initially appears.

I don't know what you're smoking, but the results show Xeon 6 using equal or less power than their AMD counterparts.

It's not as strong as what you're smoking it seems. Check the power consumption chart at the end. 375W average for 9754 vs 608W average for 6980P. Time to pass that bud along.

I don't know what you both are smoking. Using 62% more power is for sure hyperbole but Xeon 6 using equal or less power is also plain wrong.

Not hyperbole at all, 608W vs 375W averages => 162%, using the averages from the power consumption chart at the end.

[Anux]

Well well, that looks like a good development for Intel. Too bad they lost massive server market share to AMD and the last AMD gen provides a CPU replacement path for next gen.

I was quiet shocked when I saw the dies of Granite Rapids, they look massive. I think the only reason Intel uses chiplets is that they can't produce larger chips to make it monolithic. https://www.tomshardware.com/news/he...om-public-view 580 mm² for 48 cores, the yields must be horrible. Compare that to a Zen 4 chiplet (8 cores on 70 mm²) and they can decide on a daily basis in which product they want to put each chiplet (desktop/server).
I bet AMD still has a massive cost advantage.

Also Genoa uses TSMC N5 and Intel uses its "3" process node, when we consider the massive cache on Granite Rapids we can guess that the density is roughly the same maybe a little higher. But certainly far from N3.

With all that in mind, the efficiency of GR looks really impressive. It should be enough to not get totally stomped by Zen 5, we will see. Let's hope there will be better competition and prices in the future.

[drakonas777]

Without actually measuring 3D structures under electronic microscope it's hard to say what version of litho is on par with what version. Aside that we have only marketing positioning made by manufacturers themselves and very high level and not very accurate estimates based on product features.

Also if rumors to be believed a lot of TSMC clients want to jump directly to N2 family for premium ICs, because N3 given it's price has not demonstrated that much of benefits.

[jeisom]

I largely disagree when it comes to server. When it comes to desktop yes maybe comparing big to little is less relevant, the gulf in power consumption between the two is large and there are other factors. So intel has a smaller core and AMD has a bigger core not being compared, big deal. If anything the 9754 vs 6980P comparison is more valid because intel has double the cache which eats into the power budget, bringing the consumption of the cores on both parts closer together than it initially appears.

I still feel my statement stands. They are flawed direct comparisons to make all around on an architecture basis.

Not hyperbole at all, 608W vs 375W averages => 162%, using the averages from the power consumption chart at the end.

Regarding power usage, with some of the workloads, I'd take 62% more power for 100+% performance gain. Overall it really depends on what your workload is to measure what is worth it. Some workload really benefit the Xeons and others benefit the Epycs while some are dead even.

[jochendemuth]

Interesting discussion so far, but it seemingly doesn't look at the relevant facts.

The article is called "Intel Xeon 6980P vs. AMD EPYC Power Efficiency / Performance-Per-Watt Benchmarks", but it misses to graph a single performance per watt metric.
Let me do that with the summary data presented.
Screenshot from 2024-09-27 11-39-43.png
The Intel Xeon 6980P takes the top spot amongs Intel systems.

The top three spots in efficiency (performance/W) go to AMD, though. The geometric mean for the Xeon 6766E and Xeon 6780E are missing in the summary, although they mostly exist in the individual tests.
The methodology may be simplistic, but the results speak for themselves.

[bug77]

This may be just me, but when it win in performance, but it loses in perf/W, a graph showing total energy/task would be in order.

[jochendemuth]

This may be just me, but when it win in performance, but it loses in perf/W, a graph showing total energy/task would be in order.

Take the same data and graph it differently to identify where each system plots relative to performance and power consumption
Screenshot from 2024-09-27 12-13-52.png
The performance winners are top, the lowest power consumption on the left, the performance/w winners in the top left.

[bug77]

Take the same data and graph it differently to identify where each system plots relative to performance and power consumption
Screenshot from 2024-09-27 12-13-52.png
The performance winners are top, the lowest power consumption on the left, the performance/w winners in the top left.

Still, if I was building a server, I'd be interested in which CPU draws less energy overall to finish a given task.

[HEL88]

Still, if I was building a server, I'd be interested in which CPU draws less energy overall to finish a given task.

Because it is clear that you see the servers only in pictures.

Speed of execution of tasks is not important? How many simultaneous users can be served? How much space does it occupy???

The best in perf/power is EPYC 8534PN (64 Cores, TDP 175W). In terms of efficiency, it destroys EPYC 9684X, EPYC 9654, EPYC 9754 etc.

Why is it not so popular?

Because to perform the same task in time, to handle as many users as on faster processors, you need to buy 50% more of them.