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

Page 259 - 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

carrotmania

Member
Oct 3, 2020
62
171
76
🤷‍♀️ If you want to believe that Intel is hopelessly behind and will never catch up, ok. The psychological bias in seeing current realities as immutable exists, it affects us all, I feel it too, but probably worth defending against with reason. What makes you certain that this is the time that AMD becomes immune to competition, and Intel unable to adapt?

Calm down, remember you're in an AMD thread, there's no point getting your knickers twisted every time someone disses your mighty Intel.
Saying that, he said "Sapphire Rapids", and yes, Intel is HOPELESSLY BEHIND with that technology. Their chipletish design does not scale nicely AT ALL. 60cores @ 350W vs 96 (or 128) @ 320W from Epyc4.

And while Genoa is currently shipping to a few customers, wide release is expected to be on track, whereas SR is nowhere to be seen. Intel even just announced another release delay till NEXT YEAR.

AMD delivered Frontier, 1.1ExaFlops @ 21MW. Pat (Himself on the back) Gelsinger had to do the same PR limitation that Intel has always done, and quickly announced that Aurora, their 2Exa machine is out... er... soon. Lets gloss over that its 4 years late already. Turns out that 2Exa will use 60MW! 37% more watts per flipflop than AMD.
Now we have more details of El Capitan, which will do a similar 2Exa, but at 40MW. Intel will be at 50% more watts per flopflip than AMD.

I'll take AMDs release performance of the last 5 years over Intels of the last 10, any day. Intel have way more to prove than AMD currently, and I don't see them proving anything, yet.
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
I think his point is that people are over-estimating how behind Intel is.
🤷‍♀️ If you want to believe that Intel is hopelessly behind and will never catch up, ok.
Look, there's a way to go about it without getting personally invested, without making it a personal discussion about, and without having to estimate anything or creating new belief in any direction:

Companies plan roadmaps to stay competitive well into the future and estimate the competitive environment in said future. They announce those roadmaps, announce launch dates, keep them or delay them. Based on that behavior one can conclude how confident a company is about the future, and how on time the execution of those plans is.

Intel keeps being publicly bullish about its long term future. But just looking at SPR (but also Arc, Aurora and other products) Intel appears to be slacking in execution of what they previously publicly promised.

This has nothing to do with being "hopelessly behind" or "never" being able to "catch up". Plans obviously change yet more accordingly to achieve competitiveness again despite setbacks. Nothing wrong with that. And the competition as well reacts accordingly to changes in the estimated competitive environment in the future.
 
Jul 27, 2020
15,759
9,821
106
Things wouldn't look so bad for Intel if AMD weren't firing on all cylinders. It's not Intel that's behind so much as it is AMD that's way too ahead.

Imagine if AMD were still stuck with their Bulldozer architecture. In that scenario, Intel missing their targets wouldn't be big news. Intel would still have older stuff performing much better than AMD's newer stuff.

Intel's main issue has been lack of proper leadership and inability to fill the leadership roles with the right people. When Lisa Su was appointed AMD CEO, I had no idea about her abilities. I assumed she would be someone like HP's Carly Fiorina. I was wrong. She's proved herself to be one of the best female CEOs in the tech sector. Reading about Intel's culture, I doubt she would have made that far inside Intel.
 
  • Like
Reactions: Tlh97

Timmah!

Golden Member
Jul 24, 2010
1,396
603
136
Not many people use a 2nd GPU. Those that will, they will get something more expensive and probably E-ATX. Then there's also PCIe extenders or riser cables. Most people are likely to leave the 2nd slot alone or at most, use it with an M.2 NVMe adapter.

I understand that most people wont use the second slot anyway, as most dont have 2 GPUs. But i presume they are aware of that and since they put it in there anyway, they likely do it for those people, who would use it. In that case, why half-a** it like this? The problem here is the distance between the slots, even if someone wanted to put something else into the bottom slot, like M2 adapter as you mention, it would be too close to the GPU above it. They just should have moved upper slot one position up, there is clearly space, its simple as that.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Things wouldn't look so bad for Intel if AMD weren't firing on all cylinders.

AMD is not perfect. They had an excellent product in Zen, and they've iterated upon it successfully. But there have been some hiccups. Let's not ignore those. Also they continue to suffer from volume shortages. They're targeting specific, lucrative markets for many reasons, one of them being that it's the best use for their limited wafer supply. AMD can't become the Intel of 2006-2016 for that reason alone.

Still, it's best not to be narrow-minded in assuming that Intel might be "catching up" with some product like Raptor Lake or Meteor Lake when one of Intel's biggest cash cows from the past two decades is now in complete disarray. That is where AMD's decidedly competent execution stands in contrast with . . . uh . . . whatever it is that Intel is doing.
 

Asterox

Golden Member
May 15, 2012
1,026
1,775
136
🤷‍♀️ If you want to believe that Intel is hopelessly behind and will never catch up, ok. The psychological bias in seeing current realities as immutable exists, it affects us all, I feel it too, but probably worth defending against with reason. What makes you certain that this is the time that AMD becomes immune to competition, and Intel unable to adapt?

You should check Top500 list, and especially TopGreen500 list.




Well it is, Intel servers/HPC market share is sinking no doubt.This is where you can see who is the leader.If you look further into the future, i don't see any detail how Intel can turn the situation to its advantage.

Get back to reality, Intel will never have same situation or advantage as from 2007-2016.World or market has changed, the situation today or in the future will be quite different.World is in big crisis, energy is expensive and so will be even more expensive in the future.

CPU Power consumption=performance, this is and will be the most important details on the scales.

This is old Intel slide, yes it is real slide of future advances in nm technology. :mask:

OMq3C3I (3).png
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
You should check Top500 list, and especially TopGreen500 list.




Well it is, Intel servers/HPC market share is sinking no doubt.This is where you can see who is the leader.If you look further into the future, i don't see any detail how Intel can turn the situation to its advantage.

Get back to reality, Intel will never have same situation or advantage as from 2007-2016.World or market has changed, the situation today or in the future will be quite different.World is in big crisis, energy is expensive and so will be even more expensive in the future.

CPU Power consumption=performance, this is and will be the most important details on the scales.

This is old Intel slide, yes it is real slide of future advances in nm technology. :mask:

View attachment 63600
The arguments being made distill to the simple meaning that Intel has a righteous God given place at the top of the table and soon, soon, they will regain it. A sort of manifest destiny thinking applied to CPUs.
 
  • Like
Reactions: Tlh97 and ftt

Mopetar

Diamond Member
Jan 31, 2011
7,797
5,899
136
The L1 needs to be left out because it's inclusive, but there is an argument to be made that the correct way to measure Ryzen caches are L3+L2. Because that represents the maximum amount of data that can actually be cached. For an 8c Zen 4 X3D, that would be 104 MB (96+8*1).

The L3 cache is shared between all cores, but the L2 is per core. Does Zen let cores access the caches that belong to other cores? That alone would make a difference for any single core programs which now only have access to 97 MB of cache.

The access times are also vastly different. The L2 cache access times are around a third of the L3 times. If an additional 32 MB of that Rita cache were L2 instead of L3 the performance difference would be staggering, but the reported number would still be the same.

There's a reason that the various caches are all listed separately. Mixing them together like that is the kind of reporting you'd only do for people who don't even understand how a CPU works.
 

Tuna-Fish

Golden Member
Mar 4, 2011
1,324
1,462
136
The L3 cache is shared between all cores, but the L2 is per core. Does Zen let cores access the caches that belong to other cores?

You can load from other L2 caches, but no data ends up there in the first place unless the core in question accesses it.

The access times are also vastly different. The L2 cache access times are around a third of the L3 times.
L2 is faster mostly because it's closer. The other L2s are obviously not as close as your own (or even as close as the L3).
 

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
You can load from other L2 caches, but no data ends up there in the first place unless the core in question accesses it.

And the memory system needs to run cache coherency checks, but that's done against tags, IIRC, where a cache line may be 'poisoned' by a shared pointer that was altered on another core. So, no writes, just mark the that line as dirty (which must mean it can be written over). Geez, it's be around 25 years since I studied this stuff!
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
Oh dear. Let's hope nobody asks OpenAI about holocaust cloaks.

The stories I could tell you about OpenAI…

Stay away from that company. Far far away. As an example of how messed up they are, they did not filter the data coming into their model. User generated stories corrupted the model with, erm, “inappropriate” (read: illegal in any sane country) content. Rather than fixing the core issue (allowing any content to be injested), they (“alledgedly”, though nothing else explains the events before and after) threatened AI dungeon, causing mass censorship. The backlash was large enough to found a startup that has seriously put the hurt on AI Dungeon, and therefore OpenAI. The startup is not only profitable, they actually now have an arguably superior AI model. I am not naming the startup and I am being vague intentionally. I wasn’t attempting to derailed the thread.

EDIT: Disclaimer: was long time paying user of
AI Dungeon. Witnessed quite a bit. Surprised a lawsuit was not filed. Not biased, supported AID and OpenAI longer than most, but if you saw what I saw…
 
Last edited:

gruffi

Member
Nov 28, 2014
35
117
106
The good run of AMD is almost over. Enjoy Zen 4. It's not that AMD is failing. It's that Intel is improving generationally at a much faster rate. I am not saying AMD is returning to the decade of darkness and nothingness in CPU's. When Intel brings out Intel 4 aka. 7nm. That will bring Intel inline with AMD for power use. AMD can't keep up with Intel's IPC from generation to generation.
Sunny/Willow Cove and Golden Cove brought some significant IPC improvements in a quite short time. But that was just a coincidence due to Intel's 10nm delays. And they didn't solve Intel's real problems, power and area inefficiency. They actually made it even worse. As long as the competition doesn't lose massive ground IPC won't win Intel anything. Intel also won't be able to improve IPC gen by gen at the same rate. After the latest RPL leaks it seems to have +-0% IPC over ADL. And while MTL seems to be just another incremental improvement of the current architecture, Zen 5 will be a grounds up design with likely much higher IPC than Zen 4. So, AMD's "good run" as far from being over. They still have a massive p-core perf/watt and perf/area lead over Intel. Which will get even bigger with Zen 4 based products. Zen 4c and 5c are more bad news for Intel. At least Gracemont is competitive in terms of perf/area right now. AMD's c-cores will put Intel's e-cores to shame.

But there's another major problem for Intel. Their execution in recent years was BAD BAD BAD! And it doesn't seem to get better. SPR originally was planned for 2021. Now it's scheduled for 2H 2023. Even against Milan(-X) it would have been very tough to compete. In 2023 SPR has to compete against Genoa and Rembrandt. Or to make it short, Intel won't be able to compete against AMD in professional markets. And I don't see any signs it will get better for Intel in the near future. I also don't see their 7nm process to solve their problems. First they have to get it working. I wouldn't be surprised if it's affected by more delays. And even if they get it working by 2H 2023/2024, at that time AMD is readying Zen 5 on 3/4nm. Which means AMD still has a process node advantage. I would have said it might look better for Intel in the mobile market. But I'm not so sure anymore when I look at Dragon Range and Phoenix.

Personally I don't think Intel has a chance of a real comeback until 2025/26 and "Royal Core" based products. For the moment being they are just too much behind to solve all their problems within one or two generations. The competition won't take the foot off the gas.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
Things wouldn't look so bad for Intel if AMD weren't firing on all cylinders. It's not Intel that's behind so much as it is AMD that's way too ahead.

Imagine if AMD were still stuck with their Bulldozer architecture. In that scenario, Intel missing their targets wouldn't be big news. Intel would still have older stuff performing much better than AMD's newer stuff.

Intel's main issue has been lack of proper leadership and inability to fill the leadership roles with the right people. When Lisa Su was appointed AMD CEO, I had no idea about her abilities. I assumed she would be someone like HP's Carly Fiorina. I was wrong. She's proved herself to be one of the best female CEOs in the tech sector. Reading about Intel's culture, I doubt she would have made that far inside Intel.

IMO their management style is also a big factor. They honestly should split off their CPU division into a small separate company and get rid of all the useless layers of middle management. The company should be given the freedom to do what they want.
 

deasd

Senior member
Dec 31, 2013
513
724
136
Genoa spec out


clock/tdp are not as good as expected, if compare Milan 7763 to Genoa 9534, both are the same 64/128 and 280wTDP

7763EPYC$ 7,890.0015 March 202164128280 W32 MiB256 MiB2.45 GHz3.5 GHz

EPYC 953464/1282.3-2.4 GHz280W

AVX512 have negative impact on clock?
 

leoneazzurro

Senior member
Jul 26, 2016
905
1,430
136
Genoa spec out


clock/tdp are not as good as expected, if compare Milan 7763 to Genoa 9534, both are the same 64/128 and 280wTDP

7763EPYC$ 7,890.0015 March 202164128280 W32 MiB256 MiB2.45 GHz3.5 GHz

EPYC 953464/1282.3-2.4 GHz280W

AVX512 have negative impact on clock?

There is no listed boost clock (which is likely to be much higher), the listed frequency is obviously the base and TDP is obviously linked to the all core boost clock which, again, it is not listed there.
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
AVX512 have negative impact on clock?
AVX512 needs more power obviously so given a fixed TDP code will naturally run at a lower frequency when using AVX512 than when using less expensive instructions.

More important than any concrete frequency and TDP info is the perf/watt ratio. That one better makes using AVX512 worth it.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
If that's real, that's not great at all.

Is it really? For the same TDP, you get PCIe 5 instead of PCIe 4, and 12 channels of DDR5 instead of 8 channels of DDR4. So you get massively more I/O throughput and likely higher boost clocks for the same power. You also get AVX-512 over what ZEN3 had. That's not necessarily a bad thing.