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!

 

[JasonLD]

I hope AMD has an in between socket for Threadripper and workstation parts with Zen 4. Using a giant Epyc IO die and a giant SP5 sized socket seems very wasteful.

Pretty sure AMD will no longer offer non-Pro Threadrippers, let alone seperate HEDT socket.

 

[desrever]

Finally seeing the "leakers" getting their redactedexposed is great.

Also Zen 4 will be great still. 15% single thread upgrade is quite good for a gen no matter if it comes from IPC or clocks. Alder lake did about 20% with a much wider core. Does mean AMD will likely need to widen their cores in Zen 5. Multithreaded gains are also looking really good.

I think Zen4D might be even better than Zen3D by combining >5ghz clocks with the power of 3D cache. Might be what I wait for for my upgrade.





esquared
Anandtech Forum Director

 

[Asterox]

Moore's Law Is Dead and Red Gaming tech claimed so.

Tom from Moore said up to 37% ST and RGT said a 15% to 20% IPC increase.

Tom is not a reliable leaker, and especially not for AMD.Tom is a tech Youtuber(YouTube earnings + Patreons), or similar can be said for RGT.

Real or trusted leakers, a couple of them mostly only post on Twitter.

Up to 37% higher IPC, it is pure shooting at geese in a thick fog=clickbait.

 

[igor_kavinski]

Zen 4 at 5.5+ GHz not able to beat 12900KS ST performance would be kind of a disaster. 13900KS will up the bar by another 5% or so. Maybe that's how it's going to be for a few years. Intel will keep its ST performance crown while AMD will continue to deliver great multi-core scalability as well as higher power efficiency.

 

[Carfax83]

The big problem is the fanboys here do not realize that Alder Lake is ahead of Zen 3 and Raptor Lake is going to be released later this year by Intel. That means Zen 4 has to punch way above it's weight considering Alder Lake is what people are comparing Zen 4 against. Raptor Lake is the real problem for Zen 4.

Basically this. This forum has generally been leaning towards AMD since the first Zen CPU, but after Zen 3 it became almost cult like among certain forum members.

I had what I thought were reasonable, informed forum members tell me that Alder Lake only had a single digit IPC lead over Zen 3, while I'm looking at the benchmarks and seeing what looks like to me a very robust core (albeit with some weaknesses) with definite double digit IPC increase over Zen 3, in addition to much better clock speed scaling at higher clocks.

Also the constant disparaging and belittling of Alder Lake was just mind boggling to me. Stuff like this just goes to show the pitfalls of fanaticism can infiltrate damn near anyone.

Anyway, I still think Zen 4 is going to be a good product (especially the AM5 platform) and in line with my previous assessment, which was behind Raptor Lake in single thread performance, and ahead in multithreaded performance and overall efficiency. However, it does appear that Zen 4 may be substantially behind Raptor Lake in single thread performance due to the latter building on top of an already formidable Alder Lake while further mitigating it's weaknesses.

This might be disastrous to some fanboys but as already noted by several members, Golden Cove is a heavily redesigned beast while Zen 4 appears to have very minor modifications (microarchitecture wise) from it's predecessor so they can only expect so much.

 

[gdansk]

I had what I thought were reasonable, informed forum members tell me that Alder Lake only had a single digit IPC lead over Zen 3

Almost like they were going from Anandtech's SPEC2006 or SPEC2017 1T results...

 

[Hans Gruber]

Basically this. This forum has generally been leaning towards AMD since the first Zen CPU, but after Zen 3 it became almost cult like among certain forum members.

I had what I thought were reasonable, informed forum members tell me that Alder Lake only had a single digit IPC lead over Zen 3, while I'm looking at the benchmarks and seeing what looks like to me a very robust core (albeit with some weaknesses) with definite double digit IPC increase over Zen 3, in addition to much better clock speed scaling at higher clocks.

Also the constant disparaging and belittling of Alder Lake was just mind boggling to me. Stuff like this just goes to show the pitfalls of fanaticism can infiltrate damn near anyone.

Anyway, I still think Zen 4 is going to be a good product (especially the AM5 platform) and in line with my previous assessment, which was behind Raptor Lake in single thread performance, and ahead in multithreaded performance and overall efficiency. However, it does appear that Zen 4 may be substantially behind Raptor Lake in single thread performance due to the latter building on top of an already formidable Alder Lake while further mitigating it's weaknesses.

This might be disastrous to some fanboys but as already noted by several members, Golden Cove is a heavily redesigned beast while Zen 4 appears to have very minor modifications (microarchitecture wise) from it's predecessor so they can only expect so much.

I have made the point in this thread (months ago) that AMD needs to have Zen4 out yesterday which is ASAP. Zen 4 will again be the performance leader, but that lead could be gone the day Raptor Lake is released. The more time Zen4 is out to market before Raptor Lake, the more time AMD would have as #1. There are fanboys here who think that Intel has Raptor Lake finished and waiting in the wings for it's release date. That is not true, they modify and adjust the architecture based on market conditions. Meaning, what does AMD have.

The most troubling fact for AMD is the massive gains Alder Lake made in overall performance vs. Rocket Lake the previous generation. If Zen 4 is released within a few weeks to a month or two of Raptor Lake. That means even if the processors are tied or close to it. Their process lead is the only thing keeping them ahead of Intel. The next Intel CPU after Raptor is on 7nm vs. 10nm of Alder Lake and Raptor Lake. One could argue the Alder Lake improvements were benefitted from the new 10nm silicon. The same has been argued here for Zen4 on 5nm.

AMD cannot use Covid as an excuse because Alder Lake was introduced during Covid. Zen 3 was introduced during Covid. AMD let their release dates slide when insiders said that Zen 4 was ready by the end of 2021.

The concern for AMD fanboys should be the leaps and bounds that Intel is making with their CPU architecture. If Zen 5 is very good and intel keeps launching next generation chips with huge gains. All AMD will have is power efficiency, energy use and core counts. It's not like back in the days of Core2Duo. The problem is TSMC is having problems with 3nm. Does that mean that Zen 5 will be on an advanced 5nm process instead of 3nm? If so that means that Intel will make a process node gain of their own.

Things do not look bad for AMD this year or much of next year. The server side is still golden for AMD.

 

[Vattila]

I have to wonder, if Zen 4 is really just Zen 3 with 2x L2, 1.5x L2 BTB, and AVX512 on TSMC N5, then it really does seem like the vast majority of engineering resources was allocated towards developing the new platforms.

It is about risk management, I guess, and AMD CEO Lisa Su seems to be particularly good at it, as she has successfully established AMD as a high-performance leader and reliable roadmap executor — starting from near bankruptcy less than a decade ago.

Risk factors multiply, and you don't need many of them before failure becomes overwhelmingly likely. Lisa Su has made some big bets which have played out tremendously well ("Zen" roadmap, 7nm transition and Xilinx acquisition, in particular, but also PCIe 4, advanced packaging/V-Cache, and more), but she has at the same time communicated a cautious attitude, both to business growth forecasts and product performance targets. On the latter, she apparently prefers subjective description ("Zen 4 looks very good") rather than promising hard numbers. If she sets any quantifiable target, she makes sure they can beat it, even with unforeseen stumbling blocks coming their way.

Notably, "Zen 3" was a core redesign, which carries lots of risk, but it was planned in a stable platform. Very little changed except for the core. "Zen 4" uses a new high-performance optimised process (codeveloped with TSMC), and a completely new socket and platform, which all carry lots of risk, so it makes sense that they have kept the core design and topology pretty much stable. With AMD so far ahead on performance-per-watt, they had an opportunity to catch up to Intel on frequency. There is probably still a lot of software, infrastructure, customer expectation and mindshare for Intel's design choices. Matching Intel on IPC was a key goal for "Zen 3". It looks like AMD took the opportunity to aim for frequency leadership with "Zen 4".

The latter is pretty remarkable in itself. Remember, with "Zen 2" we were cautioned by AMD representatives and other industry experts that increasing frequency would be increasingly difficult on future process nodes. Many doubted that TSMC had the expertise in high-frequency process development, pointing out that their primary business was in mobile devices. With "Zen 4", AMD and TSMC have shown that they are indeed able to match Intel in this area as well, while still leading significantly in performance-per-watt. This will put AMD in a great position, especially in the server segment, as they are now increasingly competitive in a growing set of workloads.

On the desktop, where performance-per-watt has less sway, and where we (alas) will not see any core count increase, "Zen 4" may not be a knockout blow to Intel's upcoming "Raptor Lake". We may see gaming performance wars, with AMD and Intel binning for maximum frequency in the top flagship CPU models. Apart from that, AMD's not-so-secret weapon here is V-Cache.

 

[poke01]

The latter is pretty remarkable in itself. Remember, with "Zen 2" we were cautioned by AMD representatives and other industry experts that increasing frequency would be increasingly difficult on upcoming process nodes

My take is that AMD got more clocks in turn for less IPC.

Look at Apple M1 they are also on TSMC 5nm and run their M1's at 3.2Ghz. M1 has low clocks compared to AMD but have higher IPC.
AMD has the opposite, higher clocks and less IPC increase.

It's a design thing that AMD chose.

 

[majord]

Basically this. This forum has generally been leaning towards AMD since the first Zen CPU, but after Zen 3 it became almost cult like among certain forum members.

I had what I thought were reasonable, informed forum members tell me that Alder Lake only had a single digit IPC lead over Zen 3, while I'm looking at the benchmarks and seeing what looks like to me a very robust core (albeit with some weaknesses) with definite double digit IPC increase over Zen 3, in addition to much better clock speed scaling at higher clocks.

Also the constant disparaging and belittling of Alder Lake was just mind boggling to me. Stuff like this just goes to show the pitfalls of fanaticism can infiltrate damn near anyone.

Anyway, I still think Zen 4 is going to be a good product (especially the AM5 platform) and in line with my previous assessment, which was behind Raptor Lake in single thread performance, and ahead in multithreaded performance and overall efficiency. However, it does appear that Zen 4 may be substantially behind Raptor Lake in single thread performance due to the latter building on top of an already formidable Alder Lake while further mitigating it's weaknesses.

This might be disastrous to some fanboys but as already noted by several members, Golden Cove is a heavily redesigned beast while Zen 4 appears to have very minor modifications (microarchitecture wise) from it's predecessor so they can only expect so much.

Hangon. if i recall correctly, even the most enthusiastic members here keenly awaiting ADL were FAR more interested in the Gracemont cores, some even suggesting Intel should scrap the GC core linage and base everything on Gracemont, and its suposed amazing perf/watt area/watt and potential to scale to high performance..

So clearly peak ST performance was not the most exciting thing to them, but you're blaming anti ADL fanatics for ignoring these ST credentials , and now that it looks like Zen 4 may only match them at best It's suddenly ST perf is front and center? and a 'disaster' ?

Incidentally, The focus of belittling of Alderlake from some really has nothing to do with how powerful GC core is , but was squarely aimed at the overall compromises of the Halo SKU in particular at release. Overall perf/watt , The big.Little issues in win 10 , and with the Gracemont cores being limited to 8 , the need to clock them to the wall to match 5950x throughput. completely voiding the entire point of going to a Big Little architecture

 

[Racan]

Finally seeing the "leakers" getting their bullshit exposed is great.

Also Zen 4 will be great still. 15% single thread upgrade is quite good for a gen no matter if it comes from IPC or clocks. Alder lake did about 20% with a much wider core. Does mean AMD will likely need to widen their cores in Zen 5. Multithreaded gains are also looking really good.

I think that until AMD embrace the big little design philosophy it’s unrealistic to expect future Zen cores to match Intel’s P cores in IPC.

 

[deasd]

I think that until AMD embrace the big little design philosophy it’s unrealistic to expect future Zen cores to match Intel’s P cores in IPC.

not necessary to be a big+little just for 'IPC'..... in well known standard SPEC2006, SPEC2017

give 5950x '>15% ST' and would beats everything include M1/GoldenCove.

Tom is not a reliable leaker, and especially not for AMD.Tom is a tech Youtuber(YouTube earnings + Patreons), or similar can be said for RGT.

Real or trusted leakers, a couple of them mostly only post on Twitter.

Up to 37% higher IPC, it is pure shooting at geese in a thick fog=clickbait.

IIRC, MLID was 90% right when came to leak, except this time,,,
but OTOH, '+37%', '40%' IPC is not from him though, just a made up by some extreme anonymous after yesterday LisaSu's Zen4 debut. What he said is just 15-24%

 

[poke01]

but OTOH, '+37%', '40%' IPC is not from him though, just a made up by some extreme anonymous after yesterday LisaSu's Zen4 debut. What he said is just 15-24%

MLID said up to 37% ST and 15%-24% IPC. He was wrong.

give 5950x '>15% ST' and would beats everything include M1/GoldenCove.

IPC is not good. It may be beat it but thats only cause of the higher clocks.

 

[deasd]

MLID said up to 37% ST and 15%-24% IPC. He was wrong.

I'm not saying MLID's '37% ST' was right, but '37% IPC' or even '40% IPC ' is exaggerated misreading or made up by others.

IPC is not good. It may be beat it but thats only cause of the higher clocks.

We have to see how will '>15%' turns out to be. I'm not going to read to much into it until hard release.

 

[Carfax83]

Almost like they were going from Anandtech's SPEC2006 or SPEC2017 1T results...

I don't really want to go off on a tangent about Spec's reliability when it comes to predicting a CPU's performance in actual commercial applications but lets just say I'm not a fan.

Cinebench R20 had a 16% performance advantage for Golden Cove over Zen 3 with both clocked at 3.6ghz, and I think that was a better predictor of performance than Spec.

That said, IPC is really difficult to measure it seems and highly variable across applications.

 

[Det0x]

Zen 4 MT will be killer due to higher clocks. But the >15% ST uplift on Cinebench R23 compared to 5950X just makes it equal to 12900K.

Rememeber Intel will be doubling e-cores on desktop. AMD has nothing of value in the mid to low end

Can you tell me the ST uplift from 5800x up to the 5800x3d in Cinebench R23 ? Oh its negative ??
How is it possible for the 5800x3d to crush the regular 5800x so much in gaming workloads then ?

 

[poke01]

How is it possible for the 5800x3d to crush the regular 5800x so much in gaming workloads then ?

because it uses 3D V-cache

 

[Vattila]

I think that until AMD embrace the big little design philosophy it’s unrealistic to expect future Zen cores to match Intel’s P cores in IPC.

That may be so. Until they widen the core in "Zen 5", presumably, IPC will not increase much, apart from cache effects in certain workloads. That said, when it comes to big-little, AMD should have a broad design space opportunity with their modular chiplet architecture.

The "Bergamo" server chip using density-optimised CCD chiplets based on "Zen 4c" with twice the core count is scheduled to arrive in H1 next year. I wonder whether in 2-socket server systems, you will be able to have "Genoa" in one socket and "Bergamo" in the other, for a balance of high frequency and core count.

Likewise, with the modular chiplet architecture, I guess it would be relatively easy to swap "Zen 4c" CCD chiplets into the desktop package, perhaps for a 32-core HEDT product where frequency is less important. Perhaps, they could even do one of each, an 8-core "Zen 4" CCD plus a 16-core "Zen 4c" CCD, in the same package for a 24-core "big-little" product. But that is just pure speculation on my part. AMD has shared little about what comes after "Raphael" on the desktop.

 

[Carfax83]

The concern for AMD fanboys should be the leaps and bounds that Intel is making with their CPU architecture.

Yeah, this caught a lot of people off guard, including people like Andrei who write about such things for a living. I distinctly remember Andrei saying multiple times before the details of the Golden Cove architecture became known that increasing the decoders beyond 4 would be very problematic in x86 architectures due to the variable length of the instructions. This sort of commentary would usually be used when comparing Apple CPUs to x86.

But Intel did just that and smashed through that barrier, and Golden Cove is obviously very performant. Doubling the fetch bandwidth to 32 bytes was also another big deal so I was told.

If Zen 5 is very good and intel keeps launching next generation chips with huge gains. All AMD will have is power efficiency, energy use and core counts. It's not like back in the days of Core2Duo. The problem is TSMC is having problems with 3nm. Does that mean that Zen 5 will be on an advanced 5nm process instead of 3nm? If so that means that Intel will make a process node gain of their own.

It will be interesting to see whether AMD stays with 4 decoders in Zen 5. Eventually they will have to go wider to keep up with Intel I believe, because Golden Cove will be used as a foundation for the advancements of future Intel CPUs.

 

[poke01]

This sort of commentary would usually be used when comparing Apple CPUs to x86.

But Intel did just that and smashed through that barrier, and Golden Cove is obviously very performant. Doubling the fetch bandwidth to 32 bytes was also another big deal so I was told.

THIS! I was shocked to a 6 wide x86 chip. Most believed that 4 wide is the max for x86 chips. Intel broke that barrier.

 

[Vattila]

It will be interesting to see whether AMD stays with 4 decoders in Zen 5.

Zen 5 will be interesting, indeed. AMD Corporate Fellow and "Zen" lead Mike Clark is apparently giddy about it, and as I recall, CPU guru Jim Keller in interviews has been pretty optimistic about opportunities for more IPC gains through more decode bandwidth and larger instruction windows for out-of-order execution. No end in sight.

 

[tamz_msc]

Cinebench R20 had a 16% performance advantage for Golden Cove over Zen 3 with both clocked at 3.6ghz, and I think that was a better predictor of performance than Spec.

It isn't because Cinebench is a benchmark which spits out numbers based on a workload that nobody in the real world runs on a CPU.

 

[Carfax83]

So clearly peak ST performance was not the most exciting thing to them, but you're blaming anti ADL fanatics for ignoring these ST credentials , and now that it looks like Zen 4 may only match them at best It's suddenly ST perf is front and center? and a 'disaster' ?[

I'm not blaming anyone. I'm merely extrapolating the rise in pro AMD fanaticism on these forums (and all across the internet for that matter) and the blatant disparagement towards anything Intel for why so many people are freaking out that Zen 4 likely won't crush Intel, and may even be slower. It's like they created this alternate reality where Golden Cove barely matched Zen 3, and Intel 7 was complete and utter trash compared to TSMC's N5.

Incidentally, The focus of belittling of Alderlake from some really has nothing to do with how powerful GC core is , but was squarely aimed at the overall compromises of the Halo SKU in particular at release. Overall perf/watt , The big.Little issues in win 10 , and with the Gracemont cores being limited to 8 , the need to clock them to the wall to match 5950x throughput. completely voiding the entire point of going to a Big Little architecture

I get that, but despite the power characteristics of the P cores, it was still patently obvious that they were immensely powerful and capable and on another level compared to Zen 3.

 

[Det0x]

because it uses 3D V-cache

Yes and so what ?

Let me ask again, can you tell me the ST uplift from 5800x up to the 5800x3d in Cinebench R23 ? (really dont matter if we compare ST or MT)
My point is that Cinebench is not the be-all and end-all metric for "ST performance" as its clearly a very bad match for a "gaming workload" among other things.

So hypothetical Zen4 could only have "5% higher IPC" in Cinebench while having 20%higher IPC in other workloads thanks to the cache restructuring/optimization.
And on the other scale, the 45% faster than 12th gen in blender sound pretty good dont you think ?

What i'm saying is that its too early to judge the Zen4 core right now.. we need more performance numbers before doing so.

 

[Carfax83]

"Zen 4" uses a new high-performance optimised process (codeveloped with TSMC), and a completely new socket and platform, which all carry lots of risk, so it makes sense that they have kept the core design and topology pretty much stable. With AMD so far ahead on performance-per-watt, they had an opportunity to catch up to Intel on frequency. There is probably still a lot of software, infrastructure, customer expectation and mindshare for Intel's design choices. Matching Intel on IPC was a key goal for "Zen 3". It looks like AMD took the opportunity to aim for frequency leadership with "Zen 4".

This actually makes a lot of sense! I think a lot of end users don't realize how much foresight goes into designing CPUs, and often CPU architects have to account for all sorts of factors, i.e. most notably what process nodes will be available and how well it will integrate with their designs.

Intel's initial 10nm disaster is very revealing about that.