Linux Prepares For Next-Gen AMD CPUs With Up To 12 CCDs

[smitty3268]

Huh? You know efficiency gets worse as clock speed increases, right?

Uh, yes? I think you must have misunderstood me, because that was my whole point.

Usually server chips are clocked a bit lower than the desktop equivalents, because it's much more important for them to be more power efficient. The desktop side can afford to clock higher into the performance/efficiency power curve, so that they can get the fastest possible performance.

My point was it's not clear whether that will apply to Intel's e-cores or not. They're already clocked low on the desktop chip in order to be efficient. Will they be even lower on a future server chip? I have no idea.

[coder]

My point was it's not clear whether that will apply to Intel's e-cores or not. They're already clocked low on the desktop chip in order to be efficient. Will they be even lower on a future server chip? I have no idea.

Although this is only one benchmark, it gives us a number to work with:

The green bars are E-cores only. With 8 E-cores, we get package power of 48 W. If you divide it evenly, that's 6 W per core. However, we could also subtract off some overhead to get a better idea about how actual core power would scale. Just taking the 1-core and 8-core data points, I get a per-core power of about 4.7 W + 10.3 W of overhead.

If we scale that up to a 280 W package, which seems at the upper range of what server CPUs are using, these days, it only nets us 57 cores. Obviously, that's not the sort of core-count you'd target with a pure E-core server CPU. So, clocks would have to be scaled back by a nontrivial amount, at least during such a workload.

[mdedetrich]

Not that exciting. So apparently AMD will target for more performance-heavy cores. But this is inefficient for many reasons. Use cases that can scale with many cores are typically better off with multiple efficient cores than with fewer more performing ones, and applications that don't scale and instead benefit from few very powerful cores, don't need too many performance cores. Adding more performance cores is not as effective as Intel's strategy with Alder Lake and especially its successors. It is better to keep performance cores limited in number and just add multiple efficient cores. Intel gets 4 efficiency cores for every performance core they replace on the die, and it is not like the E-cores are that weak by themselves. This is an excellent approach and when schedulers mature will dominate the market i believe. I think we will also see Hyperthreading removed from Intel p-cores in the future, as it makes no sense when you have so many e-cores. This will simplify the p-cores and allow even more performance out of them.

So what AMD expects from this move to 12 CCDs? To move their product line to more cores per price point? To sell 12 core cpus to the (mainstream) desktop? This won't make much difference, honestly. They are going to keep sacrificing per-core performance for some more cores. It is not a bad improvement but Intel is going to eat their lunch. I think we are witnessing a repeat of the AMD64 days. AMD did dominate the market for a few years based on AMD64, but Intel came back with Core/Core 2 and AMD had nothing but just dual core AMD64s to compensate. I think we are at the stage of the original Core Duo (=Alder Lake). Alder Lake's successor is going to be the Core 2 moment (and a reminder here that that successor is what Zen 4 is going to face on the market since it won't come out any time soon).

If Intel CPU's didn't have shit power efficiency as of late, you would have a point. I mean for desktop (where power efficiency doesn't really matter) than this stance can be valid, but if we are talking about cloud/datacenter assuming close to ~100% utilization, AMD's is about twice power efficient as as Intel.

Furthermore that claim about scheduling is a big IF, because from what has been shown in testing the Intel scheduler is quite shitty ATM (at least on Windows 11) as shown by Steve from Gamers Nexus when he benchmarked Windows 11 vs 10. Even from a technical programming perspective, it would seem you would get best results from actually giving programmers explicit control on which cores to execute which code, which isn't how the scheduler is currently designed. I frankly cannot see how Intel's scheduler can "automagically" solve this problem to an degree that makes a meaningful difference because one thing is for sure, Alder lake hasn't shown this.

[qarium]

I was curious about this as well so when the first die shots were published I spent a bit of time with die size info (for scaling) and a tape measure.
If you omit L3 I think the numbers were something like 8 mm2 for Golden Cove, 4 mm2 for Zen3 and a bit over 2.5 mm2 for Gracemont. I included a ring agent in the AL numbers (1/4 of a ring agent for Gracemont) since the corresponding functionality in Zen3 also seemed to be included.
Die shot was for the large die with 8 GC and 8 Gracemont... GPU and display were about 30% of the die.
This is going from memory though... I had the numbers written on a piece of junk mail but cleaned up the office last weekend... might have to measure it again this weekend just to be sure.

this means according to your numbers intel use 84mm² for a 16core cpu and amd only 64mm²
for me this looks more like a marketing stunt so intel can claim they sell 16core cpus to and at the same time they claim to be the fastest in single-threat performance.
but the benchmarks show even in perfect multicore benchmarks the 8 small cores only have like 10% addition to the overall performance.
and right now these cpus perform the worst on linux... right now the linux users should better buy a AMD Ryzen 9 5950X

the prices in german(Geizhals.at):
AMD 5950X: 719€
Intel Core i9-12900KF: 599€

its 32 threats vs 24 threats...

can you tell us if we can expect price drop in the near future ?

edit: https://cpu.userbenchmark.com/Compar...m1685583vs4086

maybe AMD should consider a price drop.

the 5950X wins in this part:

64-Core OC Multi Core Mixed Speed

3330 Pts

3474 Pts

+4%

[bridgman]

this means according to your numbers intel use 84mm² for a 16core cpu and amd only 64mm²

Omitting L3 and keeping in mind the approximation of my measurements yes, I guess so... I was actually surprised by your numbers and wondered if you had a typo but they seem to check out OK. Our L3 is bigger than Intel's L3 though so it balances out a bit.

I really should redo the measurements with something more precise than the tape measure I use for construction work.

for me this looks more like a marketing stunt so intel can claim they sell 16core cpus to and at the same time they claim to be the fastest in single-threat performance.
but the benchmarks show even in perfect multicore benchmarks the 8 small cores only have like 10% addition to the overall performance.

10% sounds low, but that might be when the chip is power or thermally limited... but yes the goal of big/little is to allow both "size no object" single thread performance and efficient multi-thread performance. The Golden Cove core is wider and deeper than the Zen3 core so I would expect it to be both larger and faster in peak single thread, albeit with significant power draw.

In theory at least the small cores are supposed to be Skylake performance albeit single thread and small cache - they are surprisingly wide cores so if nothing else they're an interesting data point in core design.

can you tell us if we can expect price drop in the near future ?

Sorry, not my department

[coder]

In theory at least the small cores are supposed to be Skylake performance

Intel claims Skylake-equivalent IPC. However, they run at lower clock speed. They probably couldn't clock as high as Comet Lake without being re-architected to have shorter critical paths, which means more pipeline stages.

[coder]

I really should redo the measurements with something more precise than the tape measure I use for construction work.

Try this one?

https://www.reddit.com/r/intel/comme...lake_die_shot/

[qarium]

Omitting L3 and keeping in mind the approximation of my measurements yes, I guess so... I was actually surprised by your numbers and wondered if you had a typo but they seem to check out OK. Our L3 is bigger than Intel's L3 though so it balances out a bit.
I really should redo the measurements with something more precise than the tape measure I use for construction work.
10% sounds low, but that might be when the chip is power or thermally limited... but yes the goal of big/little is to allow both "size no object" single thread performance and efficient multi-thread performance. The Golden Cove core is wider and deeper than the Zen3 core so I would expect it to be both larger and faster in peak single thread, albeit with significant power draw.
In theory at least the small cores are supposed to be Skylake performance albeit single thread and small cache - they are surprisingly wide cores so if nothing else they're an interesting data point in core design.
Sorry, not my department

for me it looks like intel does something wrong...
they use more chipspace even of they claim their 10nm+++ fab has the same density as TSMC7nm
even in the fact of the bigger chip they charge 119€ less money compared to an 5950X this means their Profitability struggles.
and then they lose 64-Core threat performance benchmarks by 4%. and their performance on linux right now is not good. (for sure they will fix it in the future)

my AMD TR1920 is from a time who it was clear intel is faster in single core performance and amd is faster on multicore performance but this also shows that even if intel claims to be faster in singlecore performance because of the big.little design people like me will not buy intel (even if i would not hate intel)
yes i know amd also works on a big.little design but right now i would avoid any big-little design in the X86_64 cpu space.

i really don't know if it is technically possible but if AMD would lower the price of the 5950X by 100€ i am sure many people will not go with a intel big.little design.

[GreenReaper]

My hope is that Intel puts E-cores in something like the next installment of the Xeon D, which was originally conceived as an efficiency-focused cloud server variant. The Skylake-based 2nd generation kind of dropped the whole efficiency thing, but hopefully it'll come back with E-cores.

Yeah, I'm still using D-1521s for my sites. OVH picked it for their storage servers, but I found it perfectly capable for application duties, too - always seems to run at 2.7Ghz, even hyperthreaded it only gets to 42W (maybe if we were using AVX2 more...). When they refreshed this year, AMD got the design win with a 3600. (Intel is still at the top end... for now.)