Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[Exist50]

First, this is a Zen4/Genoa thread. Second, you continue to equate Golden cove in desktop to SR, which is a totally wrong assessment.

I was neither the first to reference the matter, nor did I exclusively focus on SPR. The same argument applies to the Genoa results.

And I can't believe I have to say this, but Golden Cove with more L2, AVX-512, and AMX is not weaker than desktop Golden Cove, much less by such an absurd ratio. Again, common sense goes a long way here.

 

[Exist50]

At a high level, yes they are different, but the underlying floating point execution units are plausibly the same. I don’t know why they wouldn’t share some implementation. They are likely to be built in a similar process, so even a highly optimized layout might me shared. They have to support a lot of different sizes and formats, so even the underlying FP unit would probably be quite large.

No, the actual ops supported, surrounding pipeline, and especially timing would be radically different. Maybe there's some shared code deep down, but at a practical level, the two will be wildly different.

 

[Markfw]

I was neither the first to reference the matter, nor did I exclusively focus on SPR. The same argument applies to the Genoa results.

And I can't believe I have to say this, but Golden Cove with more L2, AVX-512, and AMX is not weaker than desktop Golden Cove, much less by such an absurd ratio. Again, common sense goes a long way here.

You keep forgetting that Golden Cove <> SPR. Common sense ? Use come please.

 

[Exist50]

You keep forgetting that Golden Cove <> SPR. Common sense ? Use come please.

Sapphire Rapids uses Golden Cove. That should not be news to you in 2022.

 

[Markfw]

Sapphire Rapids uses Golden Cove. That should not be news to you in 2022.

Yes, but using the same core in a server vs a desktop cpu they will have totally different properties.

Do I have to explain everything to you like you are a 3 year old ?

 

[yuri69]

Let's see what the leaked pic shows.

The 300W+ Genoa is slower than Milan in R15, R20, and barely faster in R23. All that with bigger L2, 3.5GHz+ boost, N5-based power efficiency, and apparent "frequency optimization".

Does this look as a correct testing environment?

 

[Exist50]

Yes, but using the same core in a server vs a desktop cpu they will have totally different properties.

No, they do not, as shown by your pick of any generation that's come before. If anything, the properties of an individual core will be slightly improved by the server additions. Honestly, why double down on something so absurd? It never ages well.

 

[Markfw]

No, they do not, as shown by your pick of any generation that's come before. If anything, the properties of an individual core will be slightly improved by the server additions. Honestly, why double down on something so absurd? It never ages well.

Excuse me. FIRST, Power. SR will have 6 watts or less per core, MAX. 12900k for the golden cores has around 29 watts per core. That is such a large difference that the frequency will be WAY different, and thus the performance. And due to the number of cores and design (mesh, ring, etc) there will also be a lot of differences. Do I have to spell this out ??

 

[MadRat]

My household has Intel, AMD, and Apple products in it. I love that they are all competing. Its been 5 years since the last main desktop was upgraded, so all these new releases are very pleasing. When my wife does her 6-8 layers on the Cricut it takes her 5 year old Intel awhile to render. So in my opinion, whichever system makes reality take less time will be the best upgrade on the horizon. I no longer run any number crunching projects that are slow on even the old Intel, but I hate to fall too far off the curve. I could care less how an R15 compares to an R23, because it means very little except to pimple faced nerds. So for all the bickering we see about synthetic benchmarks here, I'd love to see them actually do workloads that more consumers can practically understand. Understand that I see no value in most Apple products but my wife does so they've got our money for iPads, iWatches, and Macbooks for the kids. I'm a PC and Android guy, so I prefer my money goes Intel and AMD. Make benchmarks for her to understand so I can justify my purchases. Happy spouse, happy house.

 

[Exist50]

Excuse me. FIRST, Power. SR will have 6 watts or less per core, MAX. 12900k for the golden cores has around 29 watts per core. That is such a large difference that the frequency will be WAY different, and thus the performance. And due to the number of cores and design (mesh, ring, etc) there will also be a lot of differences. Do I have to spell this out ??

You can go read any number of reviews to know how Golden Cove performs at various power limits. Just because you refuse to acknowledge it doesn't mean the data isn't known. Now enough humoring this insanity.

 

[jpiniero]

AMD EPYC 9664 could be the flagship 96-core Genoa CPU with 400W TDP - VideoCardz.com

AMD Genoa flagship SKU reportedly named EPYC 9664 Leaker YuuKi_AnS, better known for various pre-launch reviews of unreleased hardware, claims that the flagship of the upcoming AMD Genoa data-center CPU series might be called EPYC 9664. AMD EPYC Genoa CPU The leaker already revealed the list of...

videocardz.com

Here's what looks like the final top Genoa specs - 96 cores, 2.25 Base, 400 W TDP. That's a pretty big jump in TDP from Milan's 280.

 

[Markfw]

AMD EPYC 9664 could be the flagship 96-core Genoa CPU with 400W TDP - VideoCardz.com

AMD Genoa flagship SKU reportedly named EPYC 9664 Leaker YuuKi_AnS, better known for various pre-launch reviews of unreleased hardware, claims that the flagship of the upcoming AMD Genoa data-center CPU series might be called EPYC 9664. AMD EPYC Genoa CPU The leaker already revealed the list of...

videocardz.com

Here's what looks like the final top Genoa specs - 96 cores, 2.25 Base, 400 W TDP. That's a pretty big jump in TDP from Milan's 280.

50% jump in core count, less than 50% jump in tdp. (420 could be 50% more)

 

[maddie]

50% jump in core count, less than 50% jump in tdp. (420 could be 50% more)

That's terrible for 7>5 nm jump. Other factors at play.

 

[Karnak]

That's terrible for 7>5 nm jump. Other factors at play.

How do you know if it's terrible without knowing any real performance numbers so far? In the end perf/watt is what matters.

 

[maddie]

How do you know if it's terrible without knowing any real performance numbers so far? In the end perf/watt is what matters.

2.25 base clock versus 2.45 for 7763 model.

 

[Thibsie]

That's terrible for 7>5 nm jump. Other factors at play.

Without any performance data, your reaction is IMO ridiculous.

 

[Abwx]

2.25 base clock versus 2.45 for 7763 model.

Quite good since this include AVX512 as well as 8-10% higher IPC...

 

[maddie]

Quite good since this include AVX512 as well as 8-10% higher IPC...

If we assume a 50% power reduction for 7>5nm on the CPU chiplets & 14>6 nm on the IO.

We can approximate 0.5 x (2.25/2.45)^3 for an equivalent Milan 64 core design = 110W
=> 165W for a 96C design

This is 400W. AVX 512 and 8-10% IPC increases power by 140%?

Is this good?

 

[pakotlar]

Quite good since this include AVX512 as well as 8-10% higher IPC...

Do we have any indication it will run AVX512 within that TDP and those clocks?

 

[DrMrLordX]

It's more likely that Genoa will sustain higher boost clocks, hence the higher power target.

 

[Abwx]

If we assume a 50% power reduction for 7>5nm on the CPU chiplets & 14>6 nm on the IO.

We can approximate 0.5 x (2.25/2.45)^3 for an equivalent Milan 64 core design = 110W
=> 165W for a 96C design

This is 400W. AVX 512 and 8-10% IPC increases power by 140%?

Is this good?

AVX512 provide 2x the throughput, that mandate 2x the power for the executions units, add 50% more cores and you get the picture.

Do we have any indication it will run AVX512 within that TDP and those clocks?

That s likely, according to the numbers i quoted it would consume 3x the power at 7nm with AVX512, since their custom 5nm provide 0.5x the power they end with close to 50% more power than Milan (42% actually)..

Edit : in non AVX512 loads and same TDP it should boost 10% higher than Milan despite the 50% more cores.

 

[jpiniero]

AVX512 provide 2x the throughput

Depends on the implementation.

 

[Saylick]

If I am understanding this right, y'all are saying that AMD's base clock is the lowest guaranteed clock for all workloads, including AVX512, which is the opposite of Intel's approach because they report the base clock as the non-AVX clock and the clockspeed for AVX workloads is lower. If so, then the "typical" clock, i.e. non-AVX512, for Genoa should be noticeably higher than Milan's. Given AMD's advertised >25% PPW, which includes the 8-10% IPC gain, I would assume the clockspeed for the "typical" clock is something like ~2.6 GHz.

 

[deasd]

There's SPR-SP for comparison, which are AVX512 enabled

QS(8470): 52C/112T, 2.0-3.7Ghz, 350wTDP
ES2(8490): 60C/120T, 1.7-3.4Ghz, 350wTDP

https://twitter.com/x/status/1542359388219793409

these specs he posted might be correct. but the benchmark he did is another story.

Quite good since this include AVX512 as well as 8-10% higher IPC...

it makes me wonder whether the Raphael is AVX512 enabled or not. if yes, the clocks(5.5-5.7) are quite high, if the PPT power (230w) is just for the AVX512 specified situation....

 

[maddie]

AVX512 provide 2x the throughput, that mandate 2x the power for the executions units, add 50% more cores and you get the picture.



That s likely, according to the numbers i quoted it would consume 3x the power at 7nm with AVX512, since their custom 5nm provide 0.5x the power they end with close to 50% more power than Milan (42% actually)..

Edit : in non AVX512 loads and same TDP it should boost 10% higher than Milan despite the 50% more cores.

I don't get the picture.


64 C @ 2.45 on 7nm = 280W.
64 C @ 2.25 on 7nm = 220W.
64 C @ 2.25 on 5nm = 110W
96 cores on 5nm = 165W.
2X for AVX512 = 330W. ( accounting for 100% + for AVX512 and 50% for more cores. 110W>330W)

We have 400W here.

Maybe PCIe 5.0 is the culprit, but these numbers are high.