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!

 

[Abwx]

Review of Epyc HPC perfs at Phoronix :

https://phoronix.com/benchmark/result/amd-epyc-9684x-2p-genoa-x-benchmarks/geometric-mean-of-all-test-results-result-composite-ae92gxb.svgz

AMD EPYC 9684X Genoa-X Provides Incredible HPC Performance Review - Phoronix

www.phoronix.com

 

[Exist50]

Phoronix posted a Bergamo review. https://www.phoronix.com/review/amd-epyc-9754-bergamo

Impressive results, to say the least.

 

[Hitman928]

I don't think it's any surprise, but Intel is in serious trouble on the server side. Now we just have numbers to show it. 9684X is Genoa-X and 9754 is Bergamo. Xeon 8490H is Sapphire Rapids. Quite a few single or lightly threaded tests in Phoronix's test suite, so keep that in mind when looking at the overall average.

 

[Hitman928]

Also, in regards to power consumption:

Across all of these benchmarks carried out, the EPYC 9754 2P on average had a 385 Watt power draw... In comparison the EPYC 9654 2P had a 447 Watt average and the EPYC 9684X 2P had a 464 Watt average. And need we mention the Xeon Platinum 8490H 60-core processor consuming even more power with a 568 Watt average. The EPYC 9754 power consumption results surpassed my expectations in frankly not expecting Zen 4C to deliver such power efficiency improvements while still performing so well.

 

[Exist50]

I don't think it's any surprise, but Intel is in serious trouble on the server side. Now we just have numbers to show it. 9684X is Genoa-X and 9754 is Bergamo. Xeon 8490H is Sapphire Rapids. Quite a few single or lightly threaded tests in Phoronix's test suite, so keep that in mind when looking at the overall average.

View attachment 83262

I think Bergamo is the more competitively influential one, at least so long as the 3D V-cache SKUs remain low volume. Imagine a world where they launched Bergamo first last year, and Genoa as a niche, high per thread performance option later. Hell, with these results, do they really need regular Genoa at all?

 

[Hitman928]

I think Bergamo is the more competitively influential one, at least so long as the 3D V-cache SKUs remain low volume. Imagine a world where they launched Bergamo first last year, and Genoa as a niche, high per thread performance option later. Hell, with these results, do they really need regular Genoa at all?

I think AMD has distinct market segments that each (Bergamo, Genoa, Genoa-X) can address better than the others.

I think for many cloud customers, Bergamo is the absolute king and just blows about everything else out of the water. There are some HPC cloud type customers who may still prefer to go with something like Genoa or Genoa-X.

Genoa-X, while lower volume, will be a monster of a product for those that need the extra cache but not everyone has applications that need the extra cache. Plain Genoa fits in nicely for many enterprise customers and can be cheaper and more readily available than Genoa-X. There are still a lot of customers out there that don't need 128 or even 96 cores per CPU.

 

[DisEnchantment]

The new node is definitely helping a lot plus the lower clocks. Its 3W/Core for Bergamo vs 4.65W/Core for Genoa if glance over the ST runs.

 

[Hitman928]

Just to kind of illustrate my previous point, you can look at the geo mean from the two different reviews as the tests they used had more of an HPC focus for Genoa-X and more of a general focus for Bergamo.

HPC focused results:

More generalized test suite:

 

[Bigos]

View attachment 83264

The new node is definitely helping a lot plus the lower clocks. Its 3W/Core for Bergamo vs 4.65W/Core for Genoa if glance over the ST runs.

Isn't Bergamo confirmed 5nm?

This slide probably tells on which process Zen 4 can be implemented, and Phoenix is 4nm.

 

[DisEnchantment]

You are right it is 5nm

 

[Markfw]

I think Bergamo is the more competitively influential one, at least so long as the 3D V-cache SKUs remain low volume. Imagine a world where they launched Bergamo first last year, and Genoa as a niche, high per thread performance option later. Hell, with these results, do they really need regular Genoa at all?

As you know, I have 2 9554 (64 core) Genoa, and one 9654, and yes, the efficiency is wonderful. Also, yes, they need Genoa. The 9554 runs all 128 threads at full load at 3.5 ghz. For avx-512 work (primegrid is one example) per task it blows the 9654 out of the water, not to mention anything Intel. For other work, the cache makes a difference.

There is definitely need for both depending on the load.

 

[Joe NYC]

Just to kind of illustrate my previous point, you can look at the geo mean from the two different reviews as the tests they used had more of an HPC focus for Genoa-X and more of a general focus for Bergamo.

HPC focused results:

More generalized test suite:

View attachment 83265

And Intel still has ~75% of the server market....

 

[Ajay]

And Intel still has ~75% of the server market....

Well, the market is HUGE. Lots of small companies run servers well past their EOL dates. AMD EPYC series has just started hitting it's stride - unfortunately, that is a bit late in the game for the initial big buildout of large datacenters for cloud computer and content providers. My understanding is that there has been a bit of a slowdown following peak COVID timeframe. AMD has proven themselves to solution partners and buyers - that has taken time. The next couple of years should be very telling - provided AMD can produce the number of wafers needed to meet customer demand.

 

[Joe NYC]

Well, the market is HUGE. Lots of small companies run servers well past their EOL dates. AMD EPYC series has just started hitting it's stride - unfortunately, that is a bit late in the game for the initial big buildout of large datacenters for cloud computer and content providers. My understanding is that there has been a bit of a slowdown following peak COVID timeframe. AMD has proven themselves to solution partners and buyers - that has taken time. The next couple of years should be very telling - provided AMD can produce the number of wafers needed to meet customer demand.

I am not talking about the installed base. As late as last several quarters, Intel had ~ 75% of new datacenter CPU sales.

Even past the shortage, even past the point when AMD did not sell out its production.

 

[itsmydamnation]

And Intel is taking big margin hits to maintain that which is unsustainable. These things move slowly.....

 

[Ajay]

I am not talking about the installed base. As late as last several quarters, Intel had ~ 75% of new datacenter CPU sales.

Even past the shortage, even past the point when AMD did not sell out its production.

Wow, that's nuts. Do you have a link?

 

[Exist50]

I think AMD has distinct market segments that each (Bergamo, Genoa, Genoa-X) can address better than the others.

I think for many cloud customers, Bergamo is the absolute king and just blows about everything else out of the water. There are some HPC cloud type customers who may still prefer to go with something like Genoa or Genoa-X.

Genoa-X, while lower volume, will be a monster of a product for those that need the extra cache but not everyone has applications that need the extra cache. Plain Genoa fits in nicely for many enterprise customers and can be cheaper and more readily available than Genoa-X. There are still a lot of customers out there that don't need 128 or even 96 cores per CPU.

Yeah, but when you look at the volume distribution of the market, I think it's basically cloud > enterprise >> HPC. The only wrinkle right now is that cloud seems slow to adopt Zen 4c (see AWS statement), but that's likely to change over time. Aligning with those market volumes, it seems they'd be better off doing Bergamo -> Genoa -> Genoa V-Cache (translated to future cores).

 

[Hitman928]

Yeah, but when you look at the volume distribution of the market, I think it's basically cloud > enterprise >> HPC. The only wrinkle right now is that cloud seems slow to adopt Zen 4c (see AWS statement), but that's likely to change over time. Aligning with those market volumes, it seems they'd be better off doing Bergamo -> Genoa -> Genoa V-Cache (translated to future cores).

I’m not sure cloud is significantly ahead of enterprise just yet but either way, you also have to factor in that the non-dense cores are also used in desktop and mobile markets. AMD is server first, but I think it would be hard to justify delaying those markets to get the dense cores out first for server. I do think we’ll see AMD transition to getting them out simultaneously as they continue to grow, especially if they plan to follow Intel with a big little type design.

As for Amazon, I am skeptical of them giving an unbiased opinion about Bergamo.

 

[Abwx]

I am not talking about the installed base. As late as last several quarters, Intel had ~ 75% of new datacenter CPU sales.

Even past the shortage, even past the point when AMD did not sell out its production.

They manage to keep customers buying their inferior offering thanks to pure lies, promising that their future products will be much better than AMD s, and it work apparently, just read this report and take notice of Intel s deceptive discourse to their customers :

Intel Hopes to Stop Server Beating from AMD Next Year

After getting bruised in servers by AMD, Intel hopes to stop the bleeding in the server market with next year's chip offerings. The difference-making products will be Sierra Forest and Granite Rapids, which are due out in 2024, said Dave Zinsner, chief financial officer at Intel, last week at...

www.hpcwire.com

An exemple of the Intel lies to their customers :

There are a few reasons why Sierra Forest and Granite Rapids will stabilize Intel’s server offerings, Zinsner said.


“The performance of Granite [Rapids], relative to what customers need, and what the competitors will have, is going to be fantastic,” Zinsner said.

All things considered, “the reception has been quite good on Sapphire Rapids… in certain workloads that actually performs extremely well,” Zinsner said.

See the discourse.?.
Fantastic, they say, yet we all know that their SF and GR will be handily destroyed by already existing Bergamo and Genoa, let alone by AMD s Zen 5 based offering that will come next year at the same time as those fake marvels, as for existing SR 8490H we can see at Phoronix how it perform "extremely well"...

It s all about buying time with the hope of barely catching up at some point because currently they are roughly one if not two generations late.

Customers who currently buy Intel are jeopardising their businesses by a way bigger TCO and even more likely by the fact that 2 Intel s servers are required to compete with a single AMD one in perfs while being a disaster power comsumption wise, at the end of the day Intel with its deceptive discourse will bring them to bankruptcy, and i say it s so far so good, that s what incompetent people deserve.

 

[Markfw]

They manage to keep customers buying their inferior offering thanks to pure lies, promising that their future products will be much better than AMD s, and it work apparently, just read this report and take notice of Intel s deceptive discourse to their customers :

Intel Hopes to Stop Server Beating from AMD Next Year

After getting bruised in servers by AMD, Intel hopes to stop the bleeding in the server market with next year's chip offerings. The difference-making products will be Sierra Forest and Granite Rapids, which are due out in 2024, said Dave Zinsner, chief financial officer at Intel, last week at...

www.hpcwire.com

An exemple of the Intel lies to their customers :





See the discourse.?.
Fantastic, they say, yet we all know that their SF and GR will be handily destroyed by already existing Bergamo and Genoa, let alone by AMD s Zen 5 based offering that will come next year at the same time as those fake marvels, as for existing SR 8490H we can see at Phoronix how it perform "extremely well"...

It s all about buying time with the hope of barely catching up at some point because currently they are roughly one if not two generations late.

Customers who currently buy Intel are jeopardising their businesses by a way bigger TCO and even more likely to the fact that 2 Intel s servers are required to compete with a single AMD one in perfs while being a disaster power comsumption wise, at the end of the day Intel with its deceptive discourse will bring them to bankruptcy, and i say it s so far so good, that s what incompetent people deserve.

I can't believe the lies (well, yes I can, its Intel) Rome for the most part competes well, with SR, milan is a little better, and definitely wins, but Genoa just totally dominate in performance, and efficiency. Bergamo for cloud is another level above that. Genoa still dominates in many heavy compute tasks. At some point, the idiots in data center management HAVE to figure out that Intel is a bad choice.

 

[A///]

Intel's marketing department still thinks it's 2004 and people are incapable of doing research outside slides. Intel needs to be cleaned out of the dummies employed by them before anything improves.

 

[Abwx]

I can't believe the lies (well, yes I can, its Intel) Rome for the most part competes well, with SR, milan is a little better, and definitely wins, but Genoa just totally dominate in performance, and efficiency. Bergamo for cloud is another level above that. Genoa still dominates in many heavy compute tasks. At some point, the idiots in data center management HAVE to figure out that Intel is a bad choice.

Obviously lots of data center managers have not the slightest enginering background, more likely that most are just sophisticated accountants, and hence extremely easy to fool by ruthless liers from Intel sales dpt.

In the interview the intel guy go as far as saying that SF and GR will be better than what AMD will release next year, namely Zen 5, while we all know that it will be barely competitive with current AMD line up,.

If anything recents intel desktops leaks tell us that their Intel 3 process is just a badly implemented equivalent of TSMC s N5 and named 3 to fool the general public in believing that they caught up in the process race.

Those who buy this kool aid are good for loss of marketshare, huge time to market for those designing products and so on, as i said the incompetent will be forced out of their businesses and follow, if not precede, intel in their financial turmoil.

 

[H433x0n]

If anything recents intel desktops leaks tell us that their Intel 3 process is just a badly implemented equivalent of TSMC s N5 and named 3 to fool the general public in believing that they caught up in the process race.

Was there a desktop leak on Intel 3? I don’t think I’ve seen anything leaked that shows the performance (or lack thereof) of the Intel 3 process.

 

[Abwx]

Was there a desktop leak on Intel 3? I don’t think I’ve seen anything leaked that shows the performance (or lack thereof) of the Intel 3 process.

That s not the thread but you can find some infos in the link below, let s say that those numbers point to Intel 3 as being eventually Intel 5 for density and Intel 6 for process perf/watt, if we are to follow what is expected perf/watt wise from those slides then their futur server offering wont be competitive against Zen 5 update that will occur at the same date.

Intel Arrow Lake: Performance-Prognosen zur Leistung dämpfen Erwartungen

Erste angeblich interne Intel-Leistungsprognosen zur nächsten Desktop-CPU Arrow Lake zeigen mit Ausnahmen nur geringe Leistungszuwächse.

www.computerbase.de

 

[Exist50]

Was there a desktop leak on Intel 3? I don’t think I’ve seen anything leaked that shows the performance (or lack thereof) of the Intel 3 process.

It's just trolling. There've been pretty much no performance leaks for either Intel or AMD 2024 server products. Hopefully with silicon starting to make its way out, we see something soon.