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

Page 291 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Vattila

Senior member
Oct 22, 2004
799
1,351
136
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).

Untitled2.png


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! :)
 
Last edited:
  • Like
Reactions: richardllewis_01

biostud

Lifer
Feb 27, 2003
18,193
4,674
136
I'm not paying $100/core for their workstation part -- no margin loss here.
I'd buy a 24/32 core 7950ZX. Particularly if there's a motherboard out there that splits some of those pcie5 lanes into a crapton of more useful pcie4 lanes.... I'd pay decent money for more lanes and memory as well -- I just don't need workstation-level capability, and I'm not willing to overpay for stuff I'm not going to use....
But AMD must have a reason for scrapping the regular Threadripper line, and the only thing I can come up with is profitability or rather the lack of it.
 
Last edited:

maddie

Diamond Member
Jul 18, 2010
4,722
4,627
136
Just a slight off topic. I followed @SkyJuice60 aka Angstronomics after he leaked the correct zen 4 ipc uplift, awhile back. Same guy just leaked https://www.angstronomics.com/p/amds-rdna-3-graphics. RDNA 3 allegedly is all about area. That means amd is wafer constraint, meaning zen 4 is a bigger compromise than the previous zen iterations aka smaller ipc increase when die size plays a larger role.
Totally confused. How is that reasoned? AMD going for perf/area (low cost) tells me that they are prepped for high margins &/or lower prices.
 

dnavas

Senior member
Feb 25, 2017
355
190
116
But AMD must have a reason for scrapping the regular Threadripper line, and the only thing I can come up with is profitability or rather the lack of it.
Hanlon's Razer.
What I heard AMD say is that they talked to all of their Threadripper Pro customers and determined that what the market wants is Threadripper Pro. I'm not willing to ascribe greed or even just a profit motive to what could be plain ignorance. I can only speak for myself, a market of 1 :shrug:

Will a 32 core processor sell?
Does it make sense to create a package for it?
Does it make sense to create a package for that AND marginally higher memory and pcie bandwidth?
All different questions.

I'll wait and see what this server-lite package rumor is all about.
 
  • Like
Reactions: lightmanek

Timmah!

Golden Member
Jul 24, 2010
1,395
600
136
It's much easier to rebrand full on EPYC processors as ThreadRipper Pro(the fact that two of them can work on a 2S Motherboard confirms that) than having to keep investing on a CPU Castrated CPU brand

then again why we as customers should care about whats easier for them
if i was trx40 owner, i would be lively about TR5000s not being available for my board
 

Thibsie

Senior member
Apr 25, 2017
727
752
136
I am assuming the voltage issue with Zen 3 3D-cache will not be an issue with Zen 4. So v-cache CPU's will clock just as high as non v-cache Zen 4 chips.

It is dangerous to assume such and later complain.
There won't be as much limitation but assume there will be none is the perfect recipe for tears moaning and betrayal stories.
 
Mar 11, 2004
23,031
5,495
146
then again why we as customers should care about whats easier for them
if i was trx40 owner, i would be lively about TR5000s not being available for my board

And you just hit the nail on the head. Enthusiasts need to stop demanding that the market caters to them but then complain when it stops when they don't support it well enough for that to make sense for companies to do.
 

Timmah!

Golden Member
Jul 24, 2010
1,395
600
136
And you just hit the nail on the head. Enthusiasts need to stop demanding that the market caters to them but then complain when it stops when they don't support it well enough for that to make sense for companies to do.

again, why?
it always makes sense for companies to sell their stuff as expensive as they can - should we just shut up and pay up?

you can say vote with your wallet, and sure as hell i am by not giving them 2500 for lower end TR, but then again its not as if i had alternative going with Intel, who dont offer what i want either and if they eventaually will, they will most likely just copy AMDs behavior here. Duopoly sucks.
 

Hans Gruber

Platinum Member
Dec 23, 2006
2,092
1,065
136
It is dangerous to assume such and later complain.
There won't be as much limitation but assume there will be none is the perfect recipe for tears moaning and betrayal stories.
You can go through the forum history and see my predictions or me questioning things. Eg. the infinity fabric problems on B350/x370 motherboards not being able to run 3733/3800mhz stable. The latest AGESA bios update fixed all of that. You can drop a Zen 3 CPU in those old boards. Remember the AMD compatible memory sticks? I said there is no difference and so such need. I said all along it was a bios issue. It should have been fixed years ago but nobody complained because Zen and Zen+ CPU's did not support speeds up to or above 3600mhz. Zen 2 supported memory up to 3800mhz with the 1:1 ratio.

Or my suggestion that everyone should get 3600mhz DDR4 memory kits. That way they would be assured 3733/3800mhz memory speeds.

With regards to the 5800x3d CPU. The issue was the CPU voltage being shared with the v-cache and AMD's doubts that the v-cache memory could tolerate the higher voltage when the 5800x3d was overclocked. If Zen 4 has higher voltage tolerant v-cache or separate voltage regulators. Then the limits on core clock speeds will not be an issue for the 3d v-cache on Zen 4. It's a simple fix for a new CPU core.

At least with AMD we can complain. Intel doesn't care. When people tell Intel to fix issues. They say we will on the next motherboard and CPU generation. A deal with it approach. With AMD it's reminding them of 10 years of misery and rebuilding trust with DIY builders.

AMD has become a very solid and reliable computer setup in the last few years.
 

Kaluan

Senior member
Jan 4, 2022
500
1,071
96
Well we have the average IPC number and the max boost clock, so it's pretty straightforward to calculate the maximum ST uplift.

It's trickier for MT but AMD listed >35% in R23 example which is arguably the worst case scenario (due to being compute bound; see my edit above). If the 16C part can sustain 5.3Ghz boost on all threads (vs ~4.2Ghz for 5950X), coupled with average ~10% IPC you get ~38% better MT performance in things like R23 (matches well with >35% supplied by AMD). If it's memory bound, it can be much higher (think ~50%)
Stock/PPT locked 5950X typically clocks <4GHz all-core in CB R20 and R23.
 

Thibsie

Senior member
Apr 25, 2017
727
752
136
You can go through the forum history and see my predictions or me questioning things. Eg. the infinity fabric problems on B350/x370 motherboards not being able to run 3733/3800mhz stable. The latest AGESA bios update fixed all of that. You can drop a Zen 3 CPU in those old boards. Remember the AMD compatible memory sticks? I said there is no difference and so such need. I said all along it was a bios issue. It should have been fixed years ago but nobody complained because Zen and Zen+ CPU's did not support speeds up to or above 3600mhz. Zen 2 supported memory up to 3800mhz with the 1:1 ratio.

Or my suggestion that everyone should get 3600mhz DDR4 memory kits. That way they would be assured 3733/3800mhz memory speeds.

With regards to the 5800x3d CPU. The issue was the CPU voltage being shared with the v-cache and AMD's doubts that the v-cache memory could tolerate the higher voltage when the 5800x3d was overclocked. If Zen 4 has higher voltage tolerant v-cache or separate voltage regulators. Then the limits on core clock speeds will not be an issue for the 3d v-cache on Zen 4. It's a simple fix for a new CPU core.

At least with AMD we can complain. Intel doesn't care. When people tell Intel to fix issues. They say we will on the next motherboard and CPU generation. A deal with it approach. With AMD it's reminding them of 10 years of misery and rebuilding trust with DIY builders.

AMD has become a very solid and reliable computer setup in the last few years.

You post makes no sense. The only place you actually answer mine (that you quoted) is there :

With regards to the 5800x3d CPU. The issue was the CPU voltage being shared with the v-cache and AMD's doubts that the v-cache memory could tolerate the higher voltage when the 5800x3d was overclocked. If Zen 4 has higher voltage tolerant v-cache or separate voltage regulators. Then the limits on core clock speeds will not be an issue for the 3d v-cache on Zen 4. It's a simple fix for a new CPU core.

And it all hangs oin a IF you used your self. We DO NOT know the limitation, it has to be some. It should be less limitred than the 5800X3d, sure, AMd stated so, mind you.
The rest is unkown.
You almost sound like Nosta now, throwing whatever at the wall and see what sticks.
 

Hans Gruber

Platinum Member
Dec 23, 2006
2,092
1,065
136
You post makes no sense. The only place you actually answer mine (that you quoted) is there :



And it all hangs oin a IF you used your self. We DO NOT know the limitation, it has to be some. It should be less limitred than the 5800X3d, sure, AMd stated so, mind you.
The rest is unkown.
You almost sound like Nosta now, throwing whatever at the wall and see what sticks.
Feel free to throw insults around here. I notice you list a 5600x and a B-450F motherboard. Pick the benchmark and I will smoke your build. Someone needs to put your in your place.
 
  • Haha
Reactions: ZGR

dnavas

Senior member
Feb 25, 2017
355
190
116
when they don't support it well enough for that to make sense for companies to do.

I think we should remind ourselves that the issue for zen3 TR is availability. A quick scan at Zen2 TR pricing doesn't seem to indicate a problem with demand either. At EOL, you're seeing 45% premiums to MSRP. That's not a demand problem.

That said, there are likely cost constraints in enthusiast markets that make it difficult to amortize investment in a platform across the demand. Price sensitivity made the trx40 a hard sell, and limited stock and price gouging did nothing to make any of that easier. Regardless, it seems more likely that AMD is fine with the TR supply, it sees the market valuing a limited supply 32 core processor at nearly $100/core, so they're doing their best to make a platform worth that price so they can charge it themselves.

I predict that this move will look short-sighted in hindsight. It is predicated on demand continuing where it was trending in 2019, it discounts the pay to work market for the pay to play market at precisely the time when the play market is going to be dominated by cost concerns, and it eliminates the narrative that TR allowed for just as Intel is about to ship competitive multi-threaded CPUs. All the hand-wringing about a 24/32 core 7950++ to win against RaptorLake or MeteorLake or whatever wouldn't be necessary with a proper TR solution. But that prediction is worth exactly what you paid for it, so :shrug:
 

jamescox

Senior member
Nov 11, 2009
637
1,103
136
I think Intel will get quite some wins on MT because of their e-cores, but the question is how important that is to the buyers. Gamers will look at 3D cache models vs RPL, content creators 7900X and 7950X vs 13900K and the majority of those buying the lower SKUs will probably never know the difference of what CPU they have in their computer.
Zen 4, with larger L2 cache and a bunch of other improvements will likely see a boost in the performance benefit from SMT. You add on massive all-core clock speed improvements for Zen 4, and I suspect e-cores are going to not look very good.
 

MadRat

Lifer
Oct 14, 1999
11,909
229
106
I think the question is how can they develop an open architecture system that can compete with Apple's memory bandwidth in consumer products? A big L2 or L3 will not do it. Apple hides their costs in their mainboard specs, which AMD has been less aggressive. I'm not so sure AMD will sit back and watch Apple be the defacto video editing champion. If Apple suddenly muscles into hardcore gaming, it could be the knockout blow. The consumer market probably is less profitable than servers, but I cannot imagine it will be conceded by AMD. Right now they are going to push their strengths in gaming. But they best figure out how to make a leap in video editing, especially in the products that consumers use most.
 

jamescox

Senior member
Nov 11, 2009
637
1,103
136
Just a slight off topic. I followed @SkyJuice60 aka Angstronomics after he leaked the correct zen 4 ipc uplift, awhile back. Same guy just leaked https://www.angstronomics.com/p/amds-rdna-3-graphics. RDNA 3 allegedly is all about area. That means amd is wafer constraint, meaning zen 4 is a bigger compromise than the previous zen iterations aka smaller ipc increase when die size plays a larger role.
This isn’t new, so Zen 4 is unlikely to be any more of a “compromise” than previous Zen processors, whatever you are trying to imply there. Zen 2, with the IO die made at GF, was all about maximizing their bleeding edge wafer use. They probably sell a lot of 4 cpu chiplet Epyc processors, up to 32 core; that is only ~300 mm2 of silicon on the bleeding edge process. Compare that with intel sapphire rapids with something like 4 x 400 mm2 die for a total of ~ 1600 mm2 on their bleeding edge process. Intel might get away with that since they own their own fab, but it is an incredibly wasteful design. AMD would not be able to offer the low prices that they do without their optimization of wafer use. The 8 chiplet devices are ~600 mm2 on the expensive process, but that is still a lot less than Intel’s design. They charge a lot more for the 8 chiplet devices. It is also a big advantage over more monolithic designs since each 8 core chiplet can be binned independently. Even sapphire rapids is worse in that regard since it uses 4 x 14 core tiles that also have memory controllers.

They would obviously want to take advantage of this for their GPUs also, although this is probably more for CDNA devices rather than RDNA devices. If you look at the diagram for the “crusher” node used in the frontier supercomputer, the “IO” requirements are actually quite large.



It has 8 GPU chiplets across 4 GPUs. Every one of those gpu chiplet requires 4 IFOP-style links to the other GPU in the package and another 4 IFIS-style links off package (2 to other GPUs, 1 to the cpu, and 1 for the high speed network). It also has 4 stacks of HBM. That is close to half of an Epyc IO die; 8 IFOP, 8 IFIS, 8 channel DDR4.

The next generation might be using smaller compute chiplets with only a single stack of HBM each which requires even more interconnect. They will likely replace most of those connections with silicon bridges using EFB. It is unclear whether they will have other things in the bridge die or just multiple embedded die. Since EFB uses micro-bumps, the embedded chips can be made on completely different processes or different foundries. If it is on an advanced process, then it might be plausible for the embedded die to contain cache and such. I believe the “crusher” node is cache coherent memory throughout, so having a bridge chip with serdes links and/or memory controllers and possibly cache seems reasonable. It is unclear whether the infinity cache for GPUs will be chips stacked on top like v-cache, or chips stacked under the compute die using EFB. L3 cpu cache needs the massive connectivity allowed by SoIC stacking, but infinity cache probably doesn’t.

If RDNA3 is made up of smaller chiplets, then they would likely use the same infinity cache and other chips used for CDNA. It seems plausible to connect 4 chiplets with silicon bridges with 4 stacks of HBM for the equivalent of one big gpu with 4 HBM stacks. This allows them to make the gpu or compute chiplets on 5 nm and everything else on cheaper processes. It also allows binning each individual chiplet separately. This is more important now that GPUs are approaching what was cpu like clock speeds.
 

biostud

Lifer
Feb 27, 2003
18,193
4,674
136
I think the question is how can they develop an open architecture system that can compete with Apple's memory bandwidth in consumer products? A big L2 or L3 will not do it. Apple hides their costs in their mainboard specs, which AMD has been less aggressive. I'm not so sure AMD will sit back and watch Apple be the defacto video editing champion. If Apple suddenly muscles into hardcore gaming, it could be the knockout blow. The consumer market probably is less profitable than servers, but I cannot imagine it will be conceded by AMD. Right now they are going to push their strengths in gaming. But they best figure out how to make a leap in video editing, especially in the products that consumers use most.

I don't know how the gaming market is devided between laptops and desktops, but the entry level for an Apple desktop that cannot have its video card upgraded, will be a very hard sell to any gamer that cares about his $$$.
 
  • Like
Reactions: lightmanek

Asterox

Golden Member
May 15, 2012
1,026
1,775
136
I would guess that it slots in between 7700X and 7900X and that there won't be a 7800X-non-3D (at least for DIY). Although I am still wondering the big gap between 7600 and 7700.

We can only guess(in 2023), how AMD will arrange Zen 4 processors and prices in the end.The Desktop cpu market it is obviously not AMD's priority, especially now at the beginning of the AM5 life cycle.

AM5 is launching very soon, and then during 2023 everything will slowly fall into place.

Old/very long-lived AM4 is still very desirable, and that won't change anytime soon.


On the other side, Intel 1700 socket is dead after only two CPU generations. :mask:
 

Hans Gruber

Platinum Member
Dec 23, 2006
2,092
1,065
136
We can only guess(in 2023), how AMD will arrange Zen 4 processors and prices in the end.The Desktop cpu market it is obviously not AMD's priority, especially now at the beginning of the AM5 life cycle.

AM5 is launching very soon, and then during 2023 everything will slowly fall into place.

Old/very long-lived AM4 is still very desirable, and that won't change anytime soon.


On the other side, Intel 1700 socket is dead after only two CPU generations. :mask:
CPU sales and GPU sales are in the tank. CPU sales are probably worse than GPU sales. Zen 3 is still a very capable CPU. Intel released a dire warning for the CPU industry. My take is who cares, that means cheap CPU's for us. Now watch both Intel and AMD make pricing mistakes with the new CPU's.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,478
14,434
136
CPU sales and GPU sales are in the tank. CPU sales are probably worse than GPU sales. Zen 3 is still a very capable CPU. Intel released a dire warning for the CPU industry. My take is who cares, that means cheap CPU's for us. Now watch both Intel and AMD make pricing mistakes with the new CPU's.
5950x's are holding at $550. NEW. I did pay only about $500 for one, but the other 6 were all from 700-800. Still a great CPU. I will pay $800 in a minute for the 7950x.