Speculation: Ryzen 4000 series/Zen 3

[reqq]

guys how much better you think the mem latency will be on zen 3 with two chiplets? Thinking of selling my 3900x and pick up the successor.

 

[DrMrLordX]

True, but if there is more 5nm available, don't you think AMD would want to move also?

I'm not really sure what AMD is thinking. If the leaked roadmap is true and Warhol is the 2021 CPU, AMD may be slow-stepping their way into AM5 (or whatever they call it) with DDR5. Believe me, I'm not happy with that possibility, but AMD may have other ideas . . .

Don't we want Nosta to be right for once? Zen3 chiplets on both 7+ and 5.

It's a nice thought, and if he's right, Warhol may be a more-interesting product than just Vermeer all over again (with DDR5). That's not what we saw leaked though.

guys how much better you think the mem latency will be on zen 3 with two chiplets? Thinking of selling my 3900x and pick up the successor.

It may show up worse in synthetics, but functional memory latency in actual applications may improve just due to the changes in cache topography. I would hold out for reviews before you make any decisions if you aren't already confident that you want Vermeer.

 

[soresu]

It doesn't seem like they would want to stack multiple logic die at all.

I agree it seems unlikely at this point, but I believe it will happen eventually - albeit more likely in servers where the clock frequency is never too high anyways.

I would predict a likely minimum of N2 for 2 high - though given N2 is still only Nanosheet/GAA/MBCFET I would think that stacking logic for higher frequency desktop use would be saved for whatever future node uses forksheet transistors.

Ultimately though to achieve scaling logic stacking to anything like what has been done with memory they will need to switch to an entirely new logic device/process node paradigm that suffers far less electron leakage - probably one of the many varieties of spintronic based logic currently in development in the industry.

 

[A///]

guys how much better you think the mem latency will be on zen 3 with two chiplets? Thinking of selling my 3900x and pick up the successor.

In my own search for a vague answer, comparing the 3100X to the 3300X at the same clocks should give you a rough idea of where the new processors would be without an IPC increase. 3300X gets its cores from one CCX, the 3100 gets it from two CCX. A guy on OCUK did the math, not me, but he predicted that at the same clocks, the 3300X had a 14% advantage in games. I'm not sure if we're allowed to link to competing forums, but the user name is "Humbug" and in a thread about AMD Zen 3 in their CPU section, pages 149-151.

This doesn't extrapolate to outside of games because I don't recall that much testing done between either processor. It should only serve as a very vague idea of how performance may be assuming there is 0% IPC increase from Zen 2 to Zen 3. Gaming may show a 14% increase on the 4 cores in the 3300X, and you can't quantify that data for higher core processors, nor can to use that data for other tasks. You may see 3% increase in multi core rendering, for example, not a 14% increase in all task types, again, at a 0% IPC increase.


tl;dr: Wait and see.

Believe me, I'm not happy with that possibility, but AMD may have other ideas . . .

Are you talking about costs? I can see mobo makers getting fresh with customers and introducing new boards for no reason and deprecating support to nearly nothing in almost new hardware.

 

[DrMrLordX]

Are you talking about costs?

No. Warhol (Zen3 on AM5) is a completely unnecessary/unwanted product. They should be launching Raphael in 2021, not Warhol.

 

[NostaSeronx]

No. Warhol (Zen3 on AM5) is a completely unnecessary/unwanted product. They should be launching Raphael in 2021, not Warhol.

Warhol may be a more-interesting product than just Vermeer all over again (with DDR5).

DDR5 won't be supported till 2022, though. Under that roadmap Warhol is a 2021 product and won't be DDR5.


Don't go blindly trusting roadmaps when there is evidence that roadmap is dead wrong.

Dali/Pollock(FT5) => Warhol(FT5)
Lockhart w/ GDDR6 = Van Gogh w/ GDDR6
Cezanne is after Lucienne, not after Renoir.
Rembrandt is after Mero, not after Cezanne.
Durango is the successor to Promontory/Bixby, not Van Gogh.

Matisse(Starship CCD/Bixby IOD) -> Vermeer(Genesis CCD/Bixby IOD) -> Raphael(Badami CCD/Durango IOD)
Van Gogh -> Mero -> Rembrandt -> Phoenix
Renoir -> Lucienne -> Cezanne ->
-Dali is replaced by Renoir-
Pollock (Salvaged die for FT5) -> Warhol (Die built for FT5) ->

 

[jamescox]

That's be an ingenious approach to increasing the core count further even with N7+ (or whatever node AMD now uses for Zen 3) not significantly increasing density over Zen 2/N7. And as always smaller dies also mean higher yield. Though so far there was no single (or was there?) indication that Zen 3 is actually going to increase the core count.

I don’t think we are going to get any die stacking with Zen 3; I think it is going to be same IO die as Zen 2. They are going to need to widen the interconnect with DDR5 and PCIe 5.0 in Zen 4 or 5. It seems we still don’t know if Zen 4 will be a minor upgrade or a major change again. With the “5” generation IO, an interposer will start to make a lot of sense since they could have much wider paths with lower power. The power consumption to scale the bandwidth to PCIe 5 speeds will just be too high. I believe current IFOP links are 32-bit at around 4x memory clock and IFIS are 16-bit at around 8x memory clock. If they double that again, to PCIe 5.0 speeds, then the power consumption will increase significantly. That is basically why HBM was invented. The die area for higher and higher speed GDDR was getting larger and larger while also taking huge amounts of power.

It will be much better to use an interposer where you can easily have 1024-bit link at memory clock. HBM is a 1024-bit bus and up to 256 GB/s per stack for HBM2. connecting chiplets together with silicon interposer based interconnect will be a lot lower power than trying to use serdes links between separate chips. I am thinking that moving to an interposer is a necessity for the “5” generation. Transferring data at those speeds is just going to be too much power. They will still need PCIe 5 for connecting high performance GPUs, although at some point they will almost certainly put both on one package. If they have a compact interposer with some cpu cores, they could probably fit one or two HBM based GPUs into an EPYC like package.

So, any X3D/die stacking is probably going to be Zen 4 or 5. We may see some things diverging between desktop, workstation, and server platforms though. They would want DDR5 on desktop parts but we are no where close to actually needing PCIe 5 on the desktop parts. Even the value of PCIe 4 for desktop applications is actually questionable; is there any applications where it makes much of a difference vs. PCIe 3? There might be some new use cases where PCIe 4.0 is useful. The gaming consoles seem to be designed to do some different things to take advantage of very fast storage. I kind of suspect that PCIe 5 may be an EPYC feature, so desktop parts might be left at PCIe 4.0 for a while.

I still kind of wonder if AMD is going to make a 48 or 64 MB Zen 3 die. That isn’t particularly clever though. It is just making the best use of their MCM design. It would probably make Xeons look even more “low end” than they already do.

 

[A///]

No. Warhol (Zen3 on AM5) is a completely unnecessary/unwanted product. They should be launching Raphael in 2021, not Warhol.

You have a point, but it might be a stepping stone product for people who would want to see a small increase in performance. I'm still curious just how awful the AM5 launch will be. Buggy or does AMD somehow pull a surprise and it isn't terrible? Who knows. I get what you mean though. It feels like an XT 2.0. I was going to pick up a 3600XT a week ago but my chosen mobo was sold out and I found other excuses to stave me off from investing in Zen2 when Zen3 is right around the corner.

 

[DrMrLordX]

@A///

It's hard to say (and Warhol may be a red herring/fake leak). My guess is that AMD saw Intel sell Skylake over multiple generations and then got a taste of long-tail revenue themselves with Matisse. It's made them a bit greedy. If they don't see any reason for constant innovation, they'll learn soon enough.

edit: AM5 teething problems are likely to be significant, especially when you look back at Zen1 and see how buggy AM4 could be back then (and AM4 wasn't 100% new then either; it had already launched for Bristol Ridge in 2016). So that is something significant. Still, they've done a new platform launch with a new uarch on a new process before. It didn't stop them then (entirely). Shouldn't stop them now.

 

[JoeRambo]

Even the value of PCIe 4 for desktop applications is actually questionable; is there any applications where it makes much of a difference vs. PCIe 3? There might be some new use cases where PCIe 4.0 is useful. The gaming consoles seem to be designed to do some different things to take advantage of very fast storage. I kind of suspect that PCIe 5 may be an EPYC feature, so desktop parts might be left at PCIe 4.0 for a while.

In my opinion things are going to change during this console generation and we are already seeing the signs of that change in DX12 FL12_2. Right now we have a situation, where if you run out of GPU memory, it is game over for performance, PCIE "swapping" speed and latency kills performance outright once GPU memory is overcommitted and each frame needs to move data between GPU memory and DRAM.

But this "static" model of "we need this amount of memory for what is in this scene and for what is behind the corner so we don't drop hard in performance" is changing. Things like "Sampler Feedback" allow game app to react asynchronously and use info from GPU to manage working set in GPU memory. That means bandwidth is no longer used only during level loads, but whole time and suddenly BW and latency start to matter much more and start to get visible in averages and minimums.

 

[A///]

Build it and they will come.

That's my take on core counts and PCIE specifications. You won't get jack if you stagnate all avenues of the market.

It's hard to say (and Warhol may be a red herring/fake leak). My guess is that AMD saw Intel sell Skylake over multiple generations and then got a taste of long-tail revenue themselves with Matisse. It's made them a bit greedy. If they don't see any reason for constant innovation, they'll learn soon enough.

Quite possibly. I may have posted on here that it looked too basic and clinical. It would take me less than 20 minutes to make a fake slide myself in Adobe Illustrator or a similar program like Affinity Designer. Could be a third party having fun or AMD messing with leakers. I won't mention named but Zen 2 was plagued with leakers and people who ran wild with theories. Also I wouldn't put it past AMD's staff to toy with people because it draws up incredible conversation online.

That said, if they do an XT 2.0 then I think they could be called greedy or foolish. I think the community made it understood that they were a waste of time. The minor clock bump on what is the same process feels like a weird move to use up extra silicon they needed to clear out. I'm curious if the non XT processors are still being made. Though I'm curious whether the XT will still be produced if it's not surplus silicon and they'll carry over to Zen 3, and that my 3600 finishing off the 1600AF (2700) theory wasn't correct. No idea.

edit: AM5 teething problems are likely to be significant, especially when you look back at Zen1 and see how buggy AM4 could be back then (and AM4 wasn't 100% new then either; it had already launched for Bristol Ridge in 2016). So that is something significant. Still, they've done a new platform launch with a new uarch on a new process before. It didn't stop them then (entirely). Shouldn't stop them now.

True. AMD's long mobo lifespans have their pros and cons. Though I do recall viewing the AMD job boards for the US and abroad from November and onward and I'd seen a lot of positions opening up that would directly affect those issues. AMD's been making some big moves since Zen 2's launch. No idea if it'll pay off. If AMD can limit their bugginess to just a month or less than, to two weeks post launch, they'll win a lot of gratitude from the consumer market.

 

[NTMBK]

That said, if they do an XT 2.0 then I think they could be called greedy or foolish. I think the community made it understood that they were a waste of time. The minor clock bump on what is the same process feels like a weird move to use up extra silicon they needed to clear out. I'm curious if the non XT processors are still being made. Though I'm curious whether the XT will still be produced if it's not surplus silicon and they'll carry over to Zen 3, and that my 3600 finishing off the 1600AF (2700) theory wasn't correct. No idea.

I think you're looking at this backwards. AMD need to keep producing Zen2 so that they have something to sell until Zen3 is ready. Yields have improved enough that they can bump up the clocks and offer a marginally better product for essentially zero effort. Would you rather that they continued selling solely non-XT chips for another 6 months instead?

 

[A///]

I think you're looking at this backwards. AMD need to keep producing Zen2 so that they have something to sell until Zen3 is ready. Yields have improved enough that they can bump up the clocks and offer a marginally better product for essentially zero effort. Would you rather that they continued selling solely non-XT chips for another 6 months instead?

You misunderstood what I said. The 1600AF is still up for sale in the era of Zen2. My original presumption was that the 3600 would replace the 1600AF as the value product of the last generation. The 3600 is a pretty damn good processor. I then figured that the XT line, while not making much sense, may be sold for long after and not just the 3600, to people looking for last gen value products. Given how long AMD has always made their processors, they can upgrade to Zen 3 2-3 years from now.

Based on silly math I've seen elsewhere, Zen 3 should offer a decent performance increase over Zen 2. As to whether Zen 4 or Zen 5 will offer a huge leap in performance especially with more cores... dunno! Zen 4 may be prohibitively expensive to buy into with a new DDR5 and may not be worth it initially due to low RAM speeds.

Does that make sense now? I probably wasn't as clear as I could have been earlier. Very groggy from sleeping too much and hayfever.

 

[DrMrLordX]

AMD need to keep producing Zen2 so that they have something to sell until Zen3 is ready.

That assumes AMD couldn't have Zen3 ready by now if they really wanted it. I'm sure mobo OEMs wanted them to push back the launch when AMD reversed course and announced they would make provisions for Vermeer to work on pre-X570/X550 motherboards.

 

[Ajay]

That assumes AMD couldn't have Zen3 ready by now if they really wanted it. I'm sure mobo OEMs wanted them to push back the launch when AMD reversed course and announced they would make provisions for Vermeer to work on pre-X570/X550 motherboards.

And no credible rumors of a follow on chipset. Mobo makers have preferred the prior cadence.

 

[A///]

What I don't get is whether there will be a new mobo series or not. I'm a little surprised AMD isn't offering a suggestion to mobo makers who'd make more money by introducing or changing stuff around. They got burned by Intel and their 10th gen HEDT line which is nowhere to be seen. I sometimes see the 10980XE up for sale on B&H but that's it. Not counting other etailers with their third party listings.

AMD can quit the dog and pony show anytime they want. They may not have marketshare, but they have mindshare and the better product roll out. If they wanted to increase prices by 20-30% for Vermeer, it'll still sell like hotcakes. New mobo spec that does away with the fan. Big sales. If I were strictly a gamer and relied on it to bring me money (pro streamer) there is no way in hell I'd use anything other than Intel. A streaming rig would be AMD, of course.

 

[DisEnchantment]

Zen 3 BIOSes are downloadable from a lot of OEMs, Zen3 is close.
For my X570 board it is available since last week. Same for the B550 boards.

 

[BTRY B 529th FA BN]

Full F30 release is available for the X570 Auros Xtreme "Update AGESA ComboV2 1.0.8.1 for New Gen AMD Ryzen processors support "

EDIT: oh wow, it was F30 a day ago it's back to F30a

X570 AORUS XTREME (rev. 1.0) Support | Motherboard - GIGABYTE U.S.A.

Lasting Quality from GIGABYTE.GIGABYTE Ultra Durable™ motherboards bring together a unique blend of features and technologies that offer users the absolute ...

www.gigabyte.com

 

[Kenmitch]

Zen 3 BIOSes are downloadable from a lot of OEMs, Zen3 is close.
For my X570 board it is available since last week. Same for the B550 boards.

View attachment 29032

Probably won't happen, but it would be funny if Nvidia outed Zen 3 during their reveal.

 

[eek2121]

I wish AMD would step up and do a proper reference platform. Make sure they pay a ton of attention to every single detail and release an awesome, rock solid stable, feature filled board. Do one for each chipset. Charge appropriately, don’t undercut the competition, but don’t overcharge.

 

[eek2121]

BIOS versions likely won’t be finalized prior to release. It is interesting that the BIOS dropped already for Gigabyte boards, the release has got to be close.

 

[Makaveli]

Zen 3 BIOSes are downloadable from a lot of OEMs, Zen3 is close.
For my X570 board it is available since last week. Same for the B550 boards.

View attachment 29032

This looks similar to the Asus version just no mention of New gen and yours is AGESA combo v2 1.0.8.1

PRIME X570-PRO BIOS 2606
"Improve system performance and stability
Improve Fan control function
Improve DRAM stability
Update AM4 AGESA combo V2 PI 1.0.8.0
Improve system stability
Improve DRAM performance"

 

[eek2121]

Okay AMD, Intel and NVIDIA have outed their goodies. Your turn.

 

[trexfromouterspace]

Okay AMD, Intel and NVIDIA have outed their goodies. Your turn.

Sorry you'll have to make do with a string of emojis and a statement that Zen 3 is on track for 2020

 

[DrMrLordX]

Sorry you'll have to make do with a string of emojis and a statement that Zen 3 is on track for 2020

Those emojis better than +15% higher IPC and 20% higher all-core boost clocks. Or better!