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!

 

[exquisitechar]

That's really, REALLY late for Raphael.

I agree. Rembrandt won't be good enough in CPU performance against ADL, and with Raphael in Q1 2023 Intel will be king for all of 2022 and then some. Zen 4 desktop CPUs appear to be pretty late too. Aside from the server market, the competition will be getting fierce for AMD from late 2021 onwards.

 

[coercitiv]

I don't think Zen 4 will be 2H 2022. Based on the 15 month cadence Q1 2022 seems more likely.

Zen 3 > Zen 4 transition is expected to be different, there's a Zen3+ in the middle that is supposed to launch in H2 2021.

 

[uzzi38]

I agree. Rembrandt won't be good enough in CPU performance against ADL, and with Raphael in Q1 2023 Intel will be king for all of 2022 and then some. Zen 4 desktop CPUs appear to be pretty late too. Aside from the server market, the competition will be getting fierce for AMD from late 2021 onwards.

I think Rembrandt will be quite competitive for thin and lights in the same way Tiger Lake is now - acceptable CPU performance, strong iGPU and excellent I/O capabilities. Pure multithreaded capabilities is only important to a small segment of people (me included) in the -U space at the least.

For -H I suspect Intel will probably have a definite lead though.

 

[andermans]

I don't think Zen 4 will be 2H 2022. Based on the 15 month cadence Q1 2022 seems more likely.

The question is what happens with Zen3+. If it is more like Zen+ we'll likely have Zen3+ in Q1 2022 and Zen4 in Q2 2023. Or it could be like the 3000 XT refresh in which case it meant pretty much no delay. Given all the rumors about Genoa in 2022 I doubt Zen4 really takes until 2023, but I'm not sure when to actually expect the Zen4 release.

 

[Timorous]

Zen 3 > Zen 4 transition is expected to be different, there's a Zen3+ in the middle that is supposed to launch in H2 2021.

We have no clue what Warhol actually is. Seems like it could be a product that exists so AMD still release a desktop CPU in 2021 as a spoiler for Alder lake.

I remember people thinking the 3000 XT products meant that zen3 was delayed which was not true.

I could see a Q2 launch for zen4 but I am really skeptical that it will be 2H 2022.

 

[scineram]

Warhol is a new chip codename, not XT. At least like Lucienne or Pinnacle Ridge. Maybe more.

 

[Ajay]

Zen 3 > Zen 4 transition is expected to be different, there's a Zen3+ in the middle that is supposed to launch in H2 2021.

Zen3+ makes more sense to me now than before. With Apple likely eating up all the early 5N wafers it will be tough for AMD to get enough supply to field Zen4 products b/4 H2 2022.
Curious what AMD will do. Can't be much more frequency left on the table considering how well Ryzen 5K clocks. Logic improvements + clocks for ~10% more performance?
Also, and uplift would help against RKL, at least with gamers who care about a few extra FPS.

 

[Timorous]

Warhol is a new chip codename, not XT. At least like Lucienne or Pinnacle Ridge. Maybe more.

Warhol is a new desktop package codename. For all we know it could just be the same zen3 ccd(s) with a different IO die. What I do know is that on the roadmaps AMD have released there is nothing about a zen 3+ product so until more information is released I expect AMD to release Zen 4 inline with their 15 month cadence which is Q1 2022 and possibly Q2 2022.

 

[DrMrLordX]

With Apple likely eating up all the early 5N wafers

Apple has been on N5 for awhile. Shouldn't TSMC have had ample opportunity to expand N5 production for other customers by now?

 

[Doug S]

Apple has been on N5 for awhile. Shouldn't TSMC have had ample opportunity to expand N5 production for other customers by now?

Any expansion of N5 is limited by EUV equipment. They can only expand when they get enough deliveries from ASML to fully equip another line - and most (or even all) of those deliveries would have been allocated to N5P and more recently to N3 (which begins risk production any day now, which requires a fully equipped line)

 

[Ajay]

Apple has been on N5 for awhile. Shouldn't TSMC have had ample opportunity to expand N5 production for other customers by now?

Doh! Thanks. To think, I have an iPhone 12 with an A14 and still blew that. @Doug S nailed the problem with expanding 5N and getting 3N up and running. Fortunately for TSMC, Intel being out of the leading edge ATM has allowed TSMC to get a larger allocation.

 

[moinmoin]

Apple has been on N5 for awhile. Shouldn't TSMC have had ample opportunity to expand N5 production for other customers by now?

TSMC N5 should be in progress of reaching 120k wpm by the end of the year. N7 supposedly is at 140-150k wpm currently.

 

[LightningZ71]

It seems to me that N6 makes more sense in mobile with the APU. Better power and thermals always helps there.

 

[moinmoin]

Mark Papermaster at the Morgan Stanley Technology, Media and Telecom Conference yesterday said, next Generation Epyc (Genoa) coming in 2022 (leveraging 5nm) and Zen 5 is on track. Interestingly, he also mentioned 6nm.

Condensed transcription over at reddit. Imo notable quotes:

"Design process is 5 year cycle. The original Zen roadmap had 3 generations."

"The manufacturing node is chosen to balance performance vs reliability vs yield. (...) we wait until the node matures. We have design co-optimization with TSMC. Zen4 will be 5nm. We're on track with Zen5."

"Demand in 2021 will continue to be high. So AMD is ramping up our overall supply chain and we are confident AMD can grow despite the tight environment."

"EYPC 3 Milan is shipping in volume next month although it has been shipping in "select count" since the end of last year."

"Going forward there will not be a homogenous x86 vanilla solution like in the past. HSA and optimized compute solutions is the way forward to address different computing workloads."

https://www.reddit.com/r/AMD_Stock/comments/lwn7sm

 

[yuri69]

So Milan is shipping next month. This means Q2, right? Do we really expect Genoa to be anytime soon in 2022? Are the SP3 and SP5 platforms supposed to coexist?

 

[jpiniero]

So Milan is shipping next month. This means Q2, right? Do we really expect Genoa to be anytime soon in 2022? Are the SP3 and SP5 platforms supposed to coexist?

It's going to coexist on some level since AMD will sell plenty of Rome and Milan even after when Genoa ships. Papermaster's HSA comments make me wonder if he will see GPU chiplets on package soon enough.

 

[moinmoin]

Are the SP3 and SP5 platforms supposed to coexist?

Do we officially know about SP5 already?

Imo if AMD wants to grow quantitatively and not just raise the ASP it better keeps the capacity of older nodes as new nodes will always be more capacity constrained at first. So I do expect the old platforms to live some longer concurrently to the new ones.

 

[uzzi38]

So Milan is shipping next month. This means Q2, right? Do we really expect Genoa to be anytime soon in 2022? Are the SP3 and SP5 platforms supposed to coexist?

Yes, yes and yes.

Milan is only being held back because AMD sees no reason to release it before Ice Lake-SP launches.

 

[soresu]

It seems to me that N6 makes more sense in mobile with the APU. Better power and thermals always helps there.

From what I have heard the main benefit of N6 over N7/N7P is area scaling.

The power and frequency benefits are supposed to be minimal.

I suspect that N6 is advantageous as it will ameliorate some of the area increase from Zen2 -> Zen3 and Vega2/3 -> RDNA2 on Rembrandt, as well as the rumoured increase from 8 CU to 12 CU.

The fact that Apple is unlikely to go backwards from N5 to N6 for any significant product parts probably has its own charms too.

 

[ThatBuzzkiller]

It's going to coexist on some level since AMD will sell plenty of Rome and Milan even after when Genoa ships. Papermaster's HSA comments make me wonder if he will see GPU chiplets on package soon enough.

I think he's talking about HSA from an implementation perspective ...

HSA doesn't necessarily need to imply unified physical memory like we see with APUs or potentially chiplets on the same modules. HSA can be implemented with unified virtual memory like we see on systems using ROCm or CUDA with NVLink. In both cases we have identical programming models with different implementations ...

3rd generation infinity fabric on CDNA2 will allow us to extend HSA with multi-node CPU/GPU unified memory as pictured below:

 

[Kaloi48]

It's going to coexist on some level since AMD will sell plenty of Rome and Milan even after when Genoa ships. Papermaster's HSA comments make me wonder if he will see GPU chiplets on package soon enough.

AMD will bring iGPUs to its mainstream desktop processors Raphael according to the leaked roadmap from MebiuW.

It has a dashed outline around Navi 2, that indicates at least some of them will have Navi 2 GPUs.

 

[Tuna-Fish]

AMD will bring iGPUs to its mainstream desktop processors Raphael according to the leaked roadmap from MebiuW.
View attachment 40534
It has a dashed outline around Navi 2, that indicates at least some of them will have Navi 2 GPUs.

Totally unbacked speculation: There will be an IOD on a more modern process (but maybe still not quite state of the art) with a small GPU on it. There will also be IODs on an older process with no GPUs. This integrated GPU is aimed at the very bottom of the market, basically to run windows desktop and do video acceleration, to allow up to 16-core Zen CPUs with a tiny apu for productivity workloads without having to include an AIB GPU in the box.

 

[DisEnchantment]

Design process is 5 year cycle. The original Zen roadmap had 3 generations."

Great scoop. Does indeed sound like Zen3 was the final leg of the original Zen design. Makes sense.
Makes Zen4 design much more intriguing.

"Going forward there will not be a homogenous x86 vanilla solution like in the past. HSA and optimized compute solutions is the way forward to address different computing workloads."

The HSA vision as presented by Rick Bergman in 2011, the Fusion tech, will finally come of age. It took long enough but it is almost here.

From what I can surmise, apparently the Alderbaran GPU has a possiblity of allowing snooping in A+A combination.

[PATCH 094/159] drm/amdgpu: set snoop bit in pde/pte entries for Aldebaran A+A

Basically, the CPU is aware that a VRAM region in the PTE could be changed by GPU and can therefore now start caching those VRAM regions, (if GPU changes it that cached region is invalidated automatically, therefore no need to SW sync or load/store operations). One more step for complete cache coherent Unifed memory HSA

 

[Thibsie]

IMO the 5nm CPUs will get 7nm IOD.
7nm CPUs will keep 14/12nm IOD.
Could be wrong though.

 

[DrMrLordX]

TSMC N5 should be in progress of reaching 120k wpm by the end of the year. N7 supposedly is at 140-150k wpm currently.

That's pretty good actually. Should be plenty of wafers for AMD. TSMC knows that AMD can move product.