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!

 

[Ajay]

That's why I put 'maintain or increase margins'. If Alder Lake ends up being a dud in gaming, it'll give AMD that much more latitude in raising prices for Zen 3D. There's only so many wafers they are going to get from TSMC.

Fair enough

 

[DisEnchantment]

I think we can consider Zen+, Zen 2 XT and Zen 3D all as plan Bs that AMD would market as a quasi new gen whenever the actual new gen takes longer.

What about Warhol.
I think it was the real gap filler, it appeared in roadmaps for quite some time.
It would be interesting to know why it was cancelled in the first place.

 

[moinmoin]

Is there any reason why Zen 4 would be significantly delayed?

I already mentioned DDR5. That aside I was only stating those being plan B options AMD keeps available, which doesn't need to imply that plan A's timing is sure to fail.

What about Warhol.
I think it was the real gap filler, it appeared in roadmaps for quite some time.
It would be interesting to know why it was cancelled in the first place.

I know nothing about Warhol. We had so many contradicting guesses and no firm hints, it could be anything. Maybe it was indeed Zen 3 with AM5 IOD (requiring DDR5) which is no longer needed if Zen 4 makes it "on time" on AM5? Who knows. ¯\_(ツ)_/¯

It's notable that AMD always stated AM4 will be supported through 2020. So AM5 relying on DDR5 but still not launched makes me think DDR5 is indeed the holdup there.

 

[yuri69]

What matters more for mindshare - value or high end performance?

Because even now Zen 3 is objectively not good for value money in the 6 and 8 core categories anyway, yet it still totally wins on the mindshare battle on what is THE desktop processor series to buy.

If Zen 3D is enough to keep the gaming performance crown and even leadership in a number of productivity applications, how will the situation be any different to right now?

Well, a 12/16c 3D SKUs will surely be smashes but Intel can make much more sense this time. I mean ADL is likely faster than the OG Zen3 and is a top-to-bottom solution. It's likely there won't be any APU, 6c, or 8c 3D SKUs.

 

[Ajay]

I hope AMD is sandbagging it on Zen4. I think the normal time from first silicon to retail is ~12-15 months*. And yet, AMD claims that Zen4 is still in 'Design'. Seems a bit far fetched to me, as the development time has already been pretty long. The Zen2 design team was done around Summer 2018, then the implementation team took over. [Not quite that simple, I know]

* Of course, it hasn't been normal times so....

 

[turtile]

I hope AMD is sandbagging it on Zen4. I think the normal time from first silicon to retail is ~12-15 months*. And yet, AMD claims that Zen4 is still in 'Design'. Seems a bit far fetched to me, as the development time has already been pretty long. The Zen2 design team was done around Summer 2018, then the implementation team took over. [Not quite that simple, I know]

* Of course, it hasn't been normal times so....

It could also be the case that AMD decided to pull in more features planned for Zen 5 into Zen 4. Since AMD is starting with a brand new platform that will span many product generations, it will make sense to make the first processor family in this platform as competitive as possible. That way, they will lock in customers for drop-in upgrades for years to come.

 

[eek2121]

Alder Lake is supposed to launch before the end of this year, while Zen 4 is more than a year away. Assuming than Alder Lake is good enough, the only emergency product that AMD has available to not lose its momentum is precisely, Cache stacked Zen 3, otherwise AMD risk to be under the shadow of Alder Lake for a whole year. Doesn't seems like a good plan to let that happen.
I suppose that we will have news about that right after Intel launches Alder Lake.

One thing you made me realize: Alder Lake can use either DDR4 or DDR5. Zen 4 will be DDR5 only, so overlapping the selling period (when Zen 4 launches late next year) will let AMD offer a somewhat cheaper platform option.

Zen 3 with or without 3D cache will still be good enough for that.



Supposing Intel actually tops AMD across their full product stack, which may not be as easy as you think, how much is it going to beat AMD by?

If it's something like 5% is AMD really living under the shadow of Alder Lake? Weren't they living under an even greater shadow with Zen, Zen+, and Zen 2 in that case? However did they survive those dark times?

Alder Lake will have faster single threaded and multithreaded performance than Zen 3 with the possible exception of the 5950X (though it may beat that chip as well). AMD absolutely needs every tool in it’s war chest. If we are lucky, Zen3D will help AMD catch up.

 

[zir_blazer]

If it's something like 5% is AMD really living under the shadow of Alder Lake? Weren't they living under an even greater shadow with Zen, Zen+, and Zen 2 in that case? However did they survive those dark times?

5% faster means losing the halo effect to push other products. I don't know how much having it may affect the sales of the entire lineup, but I expect it to account for higher than zero. May be worth keeping with a counter product just to eclipse Alder Lake launch. Or to create an even higher tiered product to move the ASP upwards (Which will eventually suck for consumers, but is good for AMD).

Is there any reason why Zen 4 would be significantly delayed? Some people are claiming that they already have samples back so unless it needs a lot of tweaking or fixes, there's no real reason they couldn't launch by this time next year.

Even some of the examples don't make a lot of sense because AMD would have had incredibly firm plans about the Zen 3 launch in place by the time the Zen 2 XT parts launched. There was only a 4 month window between those and the first Zen 3 chips being released.

There could be some external factors like motherboard manufacturers needing more time (or AMD just wanting to give them more to ensure a smoother launch) or wanting better DDR5 availability, but I don't think there's anything holding Zen 4 back since the rumors make it sound as though there aren't many serious changes that would require more time fixing bugs or other performance hiccups.

One thing you made me realize: Alder Lake can use either DDR4 or DDR5. Zen 4 will be DDR5 only, so overlapping the selling period (when Zen 4 launches late next year) will let AMD offer a somewhat cheaper platform option.

Giving the chiplet approach, I believe that pairing a Zen 4 CPU die along with an AM4 IO die may be entirely possible. If DDR5 bombs but Zen 4 is ready to be produced and TSMC node is mass production ready, I even expect an AM4 Zen 4 to be a possibility, assuming than Intel gets too far ahead and AMD needs to release yet another product ASAP while the AM5 ecosystem is not ready (Do anyone remembers the first Phenoms II models, 920 and 940, that were AM2+ only and couldn't be used on AM3?).
Also, it may STILL be possible than the IMC of the next AM5 IO die supports multiple DRAM format. Giving the fact than several (Or better said, most) of the previous generations IMCs during transitional periods were hybrids (Haswell-E and Broadwell-E both supported DDR3 besides DDR4 but was only used in some very, VERY specific models and Motherboards that are rare as hell. Skylake was DDR3/DDR4 and its direct successors should, too, albeit it didn't made sense to use DDR3 that late. Phenom II was also DDR2/DDR3. Zen was DDR4 only because DDR4 was already mature and DDR3 on its way out), chances are that support is baked in just in case.

I'm still bitter about AMD never having used the 4 10G MACs available since Zen, so yeah, these dies should be entire boxes full of unannounced surprises. Both Zen 4 on AM4, and AM5 with early Motherboards supporting DDR4 or mixed slots is possible. We have already seen things like that on previous generations, so why not now?

 

[jpiniero]

I know nothing about Warhol. We had so many contradicting guesses and no firm hints, it could be anything. Maybe it was indeed Zen 3 with AM5 IOD (requiring DDR5) which is no longer needed if Zen 4 makes it "on time" on AM5? Who knows. ¯\_(ツ)_/¯

As far as we know Rembrandt Desktop is still being released on time, sometime in Q2.

AMD needs to release yet another product ASAP while the AM5 ecosystem is not ready

Zen 3D is that product.

 

[eek2121]

5% faster means losing the halo effect to push other products. I don't know how much having it may affect the sales of the entire lineup, but I expect it to account for higher than zero. May be worth keeping with a counter product just to eclipse Alder Lake launch. Or to create an even higher tiered product to move the ASP upwards (Which will eventually suck for consumers, but is good for AMD).





Giving the chiplet approach, I believe that pairing a Zen 4 CPU die along with an AM4 IO die may be entirely possible. If DDR5 bombs but Zen 4 is ready to be produced and TSMC node is mass production ready, I even expect an AM4 Zen 4 to be a possibility, assuming than Intel gets too far ahead and AMD needs to release yet another product ASAP while the AM5 ecosystem is not ready (Do anyone remembers the first Phenoms II models, 920 and 940, that were AM2+ only and couldn't be used on AM3?).
Also, it may STILL be possible than the IMC of the next AM5 IO die supports multiple DRAM format. Giving the fact than several (Or better said, most) of the previous generations IMCs during transitional periods were hybrids (Haswell-E and Broadwell-E both supported DDR3 besides DDR4 but was only used in some very, VERY specific models and Motherboards that are rare as hell. Skylake was DDR3/DDR4 and its direct successors should, too, albeit it didn't made sense to use DDR3 that late. Phenom II was also DDR2/DDR3. Zen was DDR4 only because DDR4 was already mature and DDR3 on its way out), chances are that support is baked in just in case.

I'm still bitter about AMD never having used the 4 10G MACs available since Zen, so yeah, these dies should be entire boxes full of unannounced surprises. Both Zen 4 on AM4, and AM5 with early Motherboards supporting DDR4 or mixed slots is possible. We have already seen things like that on previous generations, so why not now?

Zen 4 will not be AM4. It is not compatible with the current IO die and they aren’t going to spend money designing a new one. Furthermore, if AMD released it on AM4 they would lose board partners.

 

[eek2121]

I hope AMD is sandbagging it on Zen4. I think the normal time from first silicon to retail is ~12-15 months*. And yet, AMD claims that Zen4 is still in 'Design'. Seems a bit far fetched to me, as the development time has already been pretty long. The Zen2 design team was done around Summer 2018, then the implementation team took over. [Not quite that simple, I know]

* Of course, it hasn't been normal times so....

Why rush when you haven’t even launched your latest product? We will get Zen3D in 2-4 months, Zen 4 12 months after, and probably another surprise 12 months after that.

 

[zir_blazer]

Zen 3D is that product.

No. That is the product that should come out this year to compete with Alder Lake. I'm talking about that if by the end of the next year the AM5 ecosystem is not ready, that AMD will need yet another stopgap product.

Zen 4 will not be AM4. It is not compatible with the current IO die and they aren’t going to spend money designing a new one. Furthermore, if AMD released it on AM4 they would lose board partners.

How do you know if it is compatible with the current IO die or not? Do you have some data that I do not?

As far that I know, the only thing between the CPU die and the IO die is the IF linking them. No reason for the next IF version to not be a backwards compatible superset of the current one. So yeah, I believe that an AM4 Zen 4 is theorically possible, as the CPU die should be self-contained. That was the whole point of the chiplet approach, mix and match.

If DDR5 fails ramping up or prices are still sky high in a year and half, we will see if I was right or not. Albeit it is more likely than an AM5 with DDR4 materializes instead of Zen 4 on AM4. But I stand by my comment about it being theorically possible by just pairing it with the current IO die.

 

[jpiniero]

No. That is the product that should come out this year to compete with Alder Lake. I'm talking about that if by the end of the next year the AM5 ecosystem is not ready, that AMD will need yet another stopgap product.

It still appears that Rembrandt on AM5 is launching as scheduled. Course we don't know if Rembrandt will get a DIY release in a timely fashion but I figure DIY will be more interested in Zen 3D anyway.

DDR5 availabilty should be fine at the end of next year.

 

[eek2121]

No. That is the product that should come out this year to compete with Alder Lake. I'm talking about that if by the end of the next year the AM5 ecosystem is not ready, that AMD will need yet another stopgap product.



How do you know if it is compatible with the current IO die or not? Do you have some data that I do not?

As far that I know, the only thing between the CPU die and the IO die is the IF linking them. No reason for the next IF version to not be a backwards compatible superset of the current one. So yeah, I believe that an AM4 Zen 4 is theorically possible, as the CPU die should be self-contained. That was the whole point of the chiplet approach, mix and match.

If DDR5 fails ramping up or prices are still sky high in a year and half, we will see if I was right or not. Albeit it is more likely than an AM5 with DDR4 materializes instead of Zen 4 on AM4. But I stand by my comment about it being theorically possible by just pairing it with the current IO die.

“AM5” as the community calls it will be ready when the appropriate product that utilizes it is. I am unsure why there is confusion TBH.

EDIT: People keep assuming Rembrandt will be a desktop APU, yet I’ve seem zero evidence of this. Is there a source for this rumor? Everything I’ve seen points to it being mobile/non-socket only.

 

[DisEnchantment]

I hope AMD is sandbagging it on Zen4. I think the normal time from first silicon to retail is ~12-15 months*. And yet, AMD claims that Zen4 is still in 'Design'. Seems a bit far fetched to me, as the development time has already been pretty long. The Zen2 design team was done around Summer 2018, then the implementation team took over. [Not quite that simple, I know]

* Of course, it hasn't been normal times so....

"In Design" for something which they have silicon in their hands is a bit of a stretch
You cannot tape out something if you have not finished implementation. And you cannot implement something if you have not designed and completed design validation.
So you can say implementation is more or less completed and waiting for the steppings and doing bug fixes.

Architecture would needed to be finished even way back to three years ago or even earlier.
IIRC in 2018 on Zen one year anniversary, Mike Clark said he was already working on Zen 5 architecture

But like @uzzi38 said, in the future we can expect at best only "On Track" for launch at an unknown date in the future, that too provided they disclose something like there is a Zen 5 in the first place.
Thinking about it, now I am not sure they will even acknowledge there is a Zen 5
There will only be noise on threads like this. Like how people are guessing what Apple's next chip will be. They don't even know the name forget architecture features

 

[uzzi38]

“AM5” as the community calls it will be ready when the appropriate product that utilizes it is. I am unsure why there is confusion TBH.

EDIT: People keep assuming Rembrandt will be a desktop APU, yet I’ve seem zero evidence of this. Is there a source for this rumor? Everything I’ve seen points to it being mobile/non-socket only.

Rembrandt is in the Gigabyte leak on AM5 supported platforms. Not by name, but it's very obviously one of the 3 chip families listed as being on the platform at this date.

 

[andermans]

EDIT: People keep assuming Rembrandt will be a desktop APU, yet I’ve seem zero evidence of this. Is there a source for this rumor? Everything I’ve seen points to it being mobile/non-socket only.

What are the things you've seen that point to it being mobile-only? I haven't seen any and in absence of that I think it would be reasonable to weakly expect AMD to continue how they did things with the preceding APUs?

 

[Kedas]

Zen3 was designed with 3D stacking in mind but we didn't know about that much later after Zen3 release.
So Zen4 may also need extra work that we will not see until later, so this may also explain the extra time needed for Zen4.

 

[moinmoin]

As far as we know Rembrandt Desktop is still being released on time, sometime in Q2.

Never said otherwise. I was talking about AMD previously stating AM4 being supported through 2020, implying that the successor should arrive in 2021. Rembrandt Desktop was always for 2022 in any case, which made me think Warhol was an AM5/DDR5 chip for 2021 which was then scrapped along an AM5 platform launch this year. But as I wrote: Who knows...

 

[DrMrLordX]

It certainly does seem that AMD is taking their sweet time getting something onto the market beyond Vermeer. And their tight-lipped nature leaves us all speculating. I agree that Warhol may be scrapped entirely. But how would we ever know? They could even repurpose the name to a new product. Zen3d may wind up being Warhol even if it wasn't Warhol to begin with.

 

[Ajay]

"In Design" for something which they have silicon in their hands is a bit of a stretch
You cannot tape out something if you have not finished implementation. And you cannot implement something if you have not designed and completed design validation.
So you can say implementation is more or less completed and waiting for the steppings and doing bug fixes.

Architecture would needed to be finished even way back to three years ago or even earlier.
IIRC in 2018 on Zen one year anniversary, Mike Clark said he was already working on Zen 5 architecture

But like @uzzi38 said, in the future we can expect at best only "On Track" for launch at an unknown date in the future, that too provided they disclose something like there is a Zen 5 in the first place.
Thinking about it, now I am not sure they will even acknowledge there is a Zen 5
There will only be noise on threads like this. Like how people are guessing what Apple's next chip will be. They don't even know the name forget architecture features

I understand - I was trying to point out that the timeline seems a bit off. Mike Clark is working early on the high level design goals and perhaps the floor plan. Then his team sets out a requirements document and finally the engineers fire up their various semiconductor CAD tools and get to work. So he is ahead of the design team (and needs to be). My supposition is that the engineers came from the Zen2 team, and have had ~3 years to design Zen4. Given Keller's development structure of smaller teams and 'rapid' design - it seems reasonable that 3 years would be enough.

I do agree that the new ‘feature' of limited transparency, compared to the past, will create more thunder than lightning.

 

[eek2121]

Rembrandt is in the Gigabyte leak on AM5 supported platforms. Not by name, but it's very obviously one of the 3 chip families listed as being on the platform at this date.

How is it obvious?

What are the things you've seen that point to it being mobile-only? I haven't seen any and in absence of that I think it would be reasonable to weakly expect AMD to continue how they did things with the preceding APUs?

References to a U and H part, but none to a G part. No references about a socketed version either.

 

[uzzi38]

How is it obvious?

References to a U and H part, but none to a G part. No references about a socketed version either.

I/O and family number (19h) match up, model number is lower than Raphael. Also one more thing I don't know the specifics about and why it's the case, just the fact that it is indeed the case.

 

[eek2121]

Looks like 3 SKUs total. Type 1 could be Rembrant/Barcelo. I say this because 1 model has USB4 and 1 doesn’t. However, this could also be OEM/Non-OEM chips or chips without GPUs vs. with. Remember Raphael will have a GPU on certain SKUS.

Types 2 and 3 are interesting. Type 2 has 28 pcie lanes, but no USB4, Type 3 has 20 pcie lanes, but with USB4.

That being said, I am not saying it won’t happen, it actually makes complete sense, but I would expect it to launch with Raphael.

 

[uzzi38]

View attachment 49300

Looks like 3 SKUs total. Type 1 could be Rembrant/Barcelo. I say this because 1 model has USB4 and 1 doesn’t. However, this could also be OEM/Non-OEM chips or chips without GPUs vs. with. Remember Raphael will have a GPU on certain SKUS.

Types 2 and 3 are interesting. Type 2 has 28 pcie lanes, but no USB4, Type 3 has 20 pcie lanes, but with USB4.

That being said, I am not saying it won’t happen, it actually makes complete sense, but I would expect it to launch with Raphael.

Right below the table it actually says that Type 2 will have OPNs with and without iGPUs.

Type 3 is Phoenix.

Also, I don't expect Rembrandt to launch on the desktop that late.