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!

 

[jamescox]

Hmm, carrying this idea forward, this could be the core of a highly converged product, very similar to an APU, but, in an EPYC-like MCM with an IO die, a pair of Zen CCDs for flow control, a pair of CDNA GPUs chips, and four HBM2E or maybe even 3 stacks to feed those GPUs. I don't think that the 8-12 channels of DDR5 ECC RDIMMS will have near enough bandwidth to directly feed the GPU die, but, if we imagine the HBM chips in the same role of the infinity cache, at 8-16GB per gpu chip, and considerably more through the IO die, maybe we have something interesting. Instead of remaining at just 2P systems with a ton of PCIe lanes to expansion cards, you have quad socket systems with sufficient IO for system connectivity.

There has been rumors of such a device for a long time. AMD has shown some infinity architecture slides with 8 GPUs almost fully connected with 6 infinity fabric links. That may not be an actual implementation, but the Genoa IO die might have 6 links on each side. That would allow placing an HBM-based GPU on one side and CPU die on the other. That would allow the gpu access to possibly 12 channel DDR5 through 6 links in addition to HBM cache.

It would be great if we get such a device. It would be very compact, very high performance, and efficient. With stacking technology, they could stack cpu cores on top of the IO die and place a gpu on either side. It makes a lot of sense, but this has been rumored for a long time, so it may still be bogus. It seems plausible for the non-stacked Genoa though.

 

[MadRat]

Seems like with IC your layout could be kidney-shaped. The IO would be on one side and all your other parts fanned out as space allows, with an emphasis on minimal trace lengths wherever it matters. The newer interconnect technologies beg for something that branches out rather than continuiing down the rectangle-only by default. While chips may need to be square, packaging does not.

 

[Joe NYC]

Some Zen 4 leaks. Greymon has a very optimistic expectation for Zen 4 Genoa launch for Q1. Elsewhere, he mentioned he expects Zen 4 desktop to be Q4 2022.

https://twitter.com/x/status/1442692556358258693

 

[xilli_fiberbit]

It's not a bit too optimistic to launch Genoa in Q1 or Q2 2022 ? Because AMD partner needs samples in hands to test but it seems it's not the case for Genoa

 

[eek2121]

It's not a bit too optimistic to launch Genoa in Q1 or Q2 2022 ? Because AMD partner needs samples in hands to test but it seems it's not the case for Genoa

Genoa is launching H2, likely Q4.

 

[Abwx]

Genoa is launching H2, likely Q4.

The same leaker state that there s Zen 4 ESs tested in Tawian since Q1, and about one year, at most, is necessary for all validations.

FI Intel s ADL ESs appeared on Geekbench in late December 2020 and is released 10-11 months later.

 

[eek2121]

The same leaker state that there s Zen 4 ESs tested in Tawian since Q1, and about one year, at most, is necessary for all validations.

FI Intel s ADL ESs appeared on Geekbench in late December 2020 and is released 10-11 months later.

Since we are getting Milan-X, it wouldn’t surprise me if Genoa ends up being 2023 TBH. The leaks indicate late 2022, however, and certain statements AMD has pushed out (such as this one) also indicate a probable 2H 2022 release.

 

[Saylick]

Genoa is going to launch mid-2022 per Charlie D. I have also heard rumors that Genoa is sampling in Q1 2022, so I expect cloud/HPC providers getting the bulk of early orders in Q2 2022 before the formal public launch in mid-2022 / Q3 2022. Zen 4 for desktop, Raphael, is going to launch AFTER Genoa, so my guess is late Q3 / early Q4 2022. RDNA3 comes after that and launches in Q4.

Credit to @uzzi38 for listening in to the Susquehanna call: https://www.reddit.com/r/AMD_Stock/comments/ob7o39/comment/h3ourb9/

Okay, I'll be updating this post with all the notes from the call as they come in:

1. Genoa delayed to mid-22 vs the formerly assumed (by Charlie) early 2022. It's a delay for technical reasons, not for marketing ones - Charlie doesn't know the details but expects it to be a bug. Could be anything really, but doesn't think it's a major issue at all.
2. Some talk about Intel's public statements regarding SPR, and how it's seemingly been pushed back a quarter as sites such as Anandtech etc have reported just in the last few days.
3. Some discussion about how companies build up stock before launching chips - this is more of a general thing that all companies follow. Charlie claims that as 10nm woes have been going on, the amount of volume they've been stockpiling before a launch has dropped progressively.
4. Launch BIOSes for SPR are scheduled to arrive in late Q2, so expect volume to come afterwards. "Q3 launch wouldn't surprise". Similar situation to ICL-SP - "launch" in March, but no availability until May.
5. Originally Intel were planning on shipping SPR with some features broken just for the sake of meeting Aurora deadlines. That's no longer happening, and Aurora deadlines are slipping instead (again). By just formally delaying it's less of a mess overall, so this is actually a good thing for their customers and their image.
6. SPR-HBM is Q4. +6 months after the main SPR launch. Might as well treat it as it's own product.
7. Emerald Rapids is a "fix" to SPR. Was a backup plan in the case of 7nm issues and Granite was delayed, alongside another called Ruby Rapids. The latter won't come to market from the looks of it (so far), but EMR looks to be on the table. Nothing world-changing performance-wise, but will be a slot-in replacement to SPR.
8. Very, very little demand for ICX. Perf/W is essentially the same as current gen 14nm parts, for anything outside of specific workloads (my own input: AVX512 workloads benefit from less downclocking). Deployments are pretty low overall. Intel claims ICL-SP has SGX, but it's not secure enough at all - it covers SEM (encrypted memory) but not SVM (for virtual machines). Overall not an answer to SVE - AMD is the only choice for secure VMs still. Intel will cover this with SPR, but they haven't with ICX.
9. Intel's claims about SPR in 2021 really just hurt them, as it provided even less of a reason to even validate for ICX. Was something of an own goal.
10. Could be out of date as these rumour are a bit old, Charlie believes Intel could hit mid-2023 for real 7nm production. It slightly closes the gap to TSMC, but not enough.
11. Granite Rapids will be both "complex and powerful". However, it seems like it's a while out from now. We still have both SPR and EMR to get through still.
12. Intel management changes in the last month was a good thing. Navin Shenoy was disliked amongst customers and blamed for issues with server delays. Charlie heard 3 names in January, Navin Shenoy the first of the three. Given that, Charlie expects there to be more very high level management departures in the coming months, with the less technical more affected.
13. Charlie says there are 16 cores on each SPR die, but we'll see up to 56c for SPR SKUs - so 2 disabled per die. However he also says it's possible that Intel could launch SKUs that go up to 64c that may not exist anywhere but paper - we'll have to see. It's a decision that can be changed. However I want to step in here: I believe the 64c Charlie is referring to here is actually Emerald Rapids. From the die shots I've seen, SPR is only up to 60 cores at a max. It's a 4x4 grid on die, but 1 tile is the IMC. However EMR is rumored to go up to 64C, so I'm assuming he just has them mixed up.
14. According to Charlie Granite is now 2024. He's basing this off a 1 year cadence, but given late Q2/early Q3 SPR and EMR to come afterwards this doesn't seem like a huge stretch.
15. "Realistically there's very little on the roadmaps right now that will stop AMD". I don't need to paraphrase that one
16. That being said, Intel seems to be moving in the correct direction, no doubt in part thanks to Pat re-joining. Had Pat never re-joined, Charlie genuinely believes that Intel's roadmaps and leadership decision process would have never been enough to catch up

 

[Joe NYC]

Genoa is going to launch mid-2022 per Charlie D. I have also heard rumors that Genoa is sampling in Q1 2022, so I expect cloud/HPC providers getting the bulk of early orders in Q2 2022 before the formal public launch in mid-2022 / Q3 2022. Zen 4 for desktop, Raphael, is going to launch AFTER Genoa, so my guess is late Q3 / early Q4 2022. RDNA3 comes after that and launches in Q4.

Credit to @uzzi38 for listening in to the Susquehanna call: https://www.reddit.com/r/AMD_Stock/comments/ob7o39/comment/h3ourb9/

Thanks for posting that link to Charlie D quotes. That was also my expectation of rough time frame for Genoa, which is why I thought Q1 seemed a little optimistic. But still close to my expectation.

Also Graymon seems to be confirming another thing I expected, that Genoa will be the launch product of Zen 4, and not Ryzen, (which was an unpopular opinion at the time).

So pretty much all of the leaks line up to what I expected (including likely 2021 release of Zen 3D).

There was something that was purely hopium on my part - a potential N6 release of Zen3, with perhaps a couple of 100 MHz. This Zen 3 XT could be a tidbit pointing in that direction.

 

[Ajay]

The same leaker state that there s Zen 4 ESs tested in Tawian since Q1, and about one year, at most, is necessary for all validations.

FI Intel s ADL ESs appeared on Geekbench in late December 2020 and is released 10-11 months later.

IIRC, it's about 18 months for server. More extensive validation RAS, in band/out of band management, more I/O, more configurations. Testing against in major, diverse, server OEM platforms, test in SAS, NAS, enterprise, cloud, VMs and so on.

 

[Ajay]

I saw some blip on AMD raising prices (above 5XXX) for Zen4 Raphael. Can't find it again, anyone else see it?

 

[Joe NYC]

I saw some blip on AMD raising prices (above 5XXX) for Zen4 Raphael. Can't find it again, anyone else see it?

I am gong to guess some YouTuber going in circles:

1. performance increase over X.
"Will it be less than X?"
"More than X?"
"I am going to be honest with you, it will be some fraction or multiple of X"

2.Date of launch
"Will it be before D?"
"After D?"
"I am going to be honest with you, it will be some time before D, but there is a chance it will be after D"

2.Price of product
"Will it be higher than X?"
"Lower than X?"
"I am going to be honest with you, it could be one of the 2, but based on the market conditions, it may reverse"

 

[moinmoin]

I saw some blip on AMD raising prices (above 5XXX) for Zen4 Raphael. Can't find it again, anyone else see it?

With Zen 3 AMD consequently increased the MSRP prices by the amount of performance increase the upgrade offered over Zen 2. AMD might want to try doing the same again with Zen 4.

 

[Joe NYC]

With Zen 3 AMD consequently increased the MSRP prices by the amount of performance increase the upgrade offered over Zen 2. AMD might want to try doing the same again with Zen 4.

I think that scenario was quite unique since Intel did not have a competitive product and there was a severe shortage.

In shortage like that, prices any lower than AMD charged would just inflate profits of scalpers, with no benefits to consumers or AMD.

I think Intel prices and Zen 4 performance vs. Intel performance will be be more relevant metrics if the shortage is behind us a year from now.

 

[lightmanek]

With Zen 3 AMD consequently increased the MSRP prices by the amount of performance increase the upgrade offered over Zen 2. AMD might want to try doing the same again with Zen 4.

Yes but no, I think the price will be adjusted according to Intel and not current AMD offering.

 

[yuri69]

It's not a bit too optimistic to launch Genoa in Q1 or Q2 2022 ? Because AMD partner needs samples in hands to test but it seems it's not the case for Genoa

Exactly, the stated "since Q1 in Taiwan" doesn't really feel true. There have been no single leak for Genoa - no CPUID leak, no benchmark DB leak, no BIOS/AGESA leak, no ES chip photo, etc. So partners do not seem to have access to it.

So this Q1 2022 target could probably mean a hyperscaler-only pseudo-launch and not real availability.

 

[moinmoin]

I think that scenario was quite unique since Intel did not have a competitive product and there was a severe shortage.

In shortage like that, prices any lower than AMD charged would just inflate profits of scalpers, with no benefits to consumers or AMD.

I think Intel prices and Zen 4 performance vs. Intel performance will be be more relevant metrics if the shortage is behind us a year from now.

Yes but no, I think the price will be adjusted according to Intel and not current AMD offering.

So far with Zen products AMD has always pursued a price to performance ratio superior to Intel's products. And that's the thing about price increases in line with performance increases: the price/performance ratio doesn't change (this also has the advantage of not unnecessarily devaluing prior products). If Intel finally offers a better price to performance ratio than AMD then we can be sure AMD will react to that at some point. But neither has happened yet.

 

[itsmydamnation]

Exactly, the stated "since Q1 in Taiwan" doesn't really feel true. There have been no single leak for Genoa - no CPUID leak, no benchmark DB leak, no BIOS/AGESA leak, no ES chip photo, etc. So partners do not seem to have access to it.

So this Q1 2022 target could probably mean a hyperscaler-only pseudo-launch and not real availability.

by that logic RDNA2 has never been released, because we never got performance leaks etc. only thing we got was info in published drivers.

it seems AMD over time has become increasingly tight on information control, absence of evidence is not evidence of absence.

 

[Saylick]

by that logic RDNA2 has never been released, because we never got performance leaks etc. only thing we got was info in published drivers.

it seems AMD over time has become increasingly tight on information control, absence of evidence is not evidence of absence.

Agreed. There's definitely progress being made and conversations happening between AMD and their partners; I mean, just look at what information got leaked when Gigabyte was hacked. There's clearly work being done in the background, but AMD obviously doesn't want it to leak.

 

[DrMrLordX]

I have also heard rumors that Genoa is sampling in Q1 2022

That seems a little late . . . Genoa ES has been out for months already.

 

[yuri69]

by that logic RDNA2 has never been released, because we never got performance leaks etc. only thing we got was info in published drivers.

it seems AMD over time has become increasingly tight on information control, absence of evidence is not evidence of absence.

Partner channels leak the info in most cases - roadmaps, prod planning, development drafts, docs, specs, reference manuals, various pre-prod samples, validation procedure checklists, etc.

Partner involvement is way bigger in case of a CPU launch. A GPU launch usually means placing a sticker on the reference SKU. A CPU launch involves designing a whole new mobo based on the reference platform.

You can't really compare a GPU (a single refernce SKU) and a CPU (custom mobos) launch.

 

[Saylick]

That seems a little late . . . Genoa ES has been out for months already.

Internal engineering samples can be different than the samples sent to partners for validation, right? When I said sampling, I meant the ones sent out to partners for validation.

Penetration Rate of Ice Lake CPUs in Server Market Expected to Surpass 30% by Year's End as x86 Architecture Remains Dominant, Says TrendForce

Genoa CPUs, AMD’s competitive equivalent of the Intel Eagle Stream, are expected to enter mass production on a similar schedule, since AMD’s wafer starts at the 5nm node have been relatively low-volume. AMD’s server processors manufactured at the 14nm node and below have the competitive advantage in terms of price-to-performance, core count, and interface support. Furthermore, after progressing to the 7nm node, these processors have been seeing gradually increased adoption by various public cloud service providers, including Google Cloud Platform, Microsoft Azure, and Tencent, throughout 2021. AMD CPUs have currently surpassed a 10% penetration rate in these three CSPs' servers. Going forward, AMD will begin inputting wafers at the 5nm node at the end of 2021 in order to further optimize its processors’ cost, power consumption, and performance. TrendForce therefore expects AMD CPUs to reach a 15% share in the global server market in 2022.

My understanding is that it is typically 3-4 quarters between the first batch of wafers to commercial availability. If H2 2022 launch is to be believed then Q3-Q4 is when AMD will have to send in their first N5 order to TSMC, which aligns with what TrendForce says. I expect this first batch to be used as validation samples for their partners, with the actual HVM ramp up being sometime in Q2 2022 since it takes ~3 months to process a wafer from beginning to end.

 

[ashFTW]

I expect this first batch to be used as validation samples for their partners, with the actual HVM ramp up being sometime in Q2 2022 since it takes ~3 months to process a wafer from beginning to end.

Doesn’t that assume that there are no issues during validation?

 

[Saylick]

Doesn’t that assume that there are no issues during validation?

Not sure, but I think they have a 2-3 month window, i.e. AMD puts in the first wafer batch order right about now, they come back end of this year/beginning of next year, spend a few months sampling to vendors, then HVM in April or so. Full launch in July/August.

 

[Ajay]

with the actual HVM ramp up being sometime in Q2 2022 since it takes ~3 months to process a wafer from beginning to end.

IIRC, it's closer to 4 months right now (actually, that was pre-COVID - so, maybe worse??).