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!

 

[Thunder 57]

Architecturally no but the 5600X3D was released in limited quantities just this year. Who knows what other scrap dies AMD may decide to unload on AM4 in future. At any rate, if AMD wants to, AMD can. They just need to open their minds more and stop chasing the dollar like a mad canine. Get back to their roots and serve the enthusiasts.

Eh, I think AMD is done with AM4. AM5 boards and CPU's have come down in price as well as DDR5. I expect them to be focusing on that. I am curious to see what Zen 5 brings.

 

[Schmide]

Eh, I think AMD is done with AM4. AM5 boards and CPU's have come down in price as well as DDR5. I expect them to be focusing on that. I am curious to see what Zen 5 brings.

May 23, 23 Dr. Lisa Su said

AM4 is a great platform that will continue for many years to come

 

[moinmoin]

Eh, I think AMD is done with AM4.

If nothing else it's already guaranteed to live on in embedded for years to come.

News - Five more years of official AM4 support

It seems this news flew under the radar at many places since the focus is on embedded. On April 20th AMD announced Ryzen Embedded 5000 series: https://www.amd.com/en/newsroom/press-releases/2023-4-20-amd-bolsters-embedded-portfolio-with-new-ryzen-emb.html Like all embedded products it has an...

forums.anandtech.com

 

[Thunder 57]

Damn you all are trying to make this into something else.

May 23, 23 Dr. Lisa Su said

Sure, it will live on for years as there isn't much reason for people on Zen 3 to upgrade anytime soon.

If nothing else it's already guaranteed to live on in embedded for years to come.

News - Five more years of official AM4 support

It seems this news flew under the radar at many places since the focus is on embedded. On April 20th AMD announced Ryzen Embedded 5000 series: https://www.amd.com/en/newsroom/press-releases/2023-4-20-amd-bolsters-embedded-portfolio-with-new-ryzen-emb.html Like all embedded products it has an...

forums.anandtech.com

I was clearly talking about socketed AM4 boards. So while that is true that is not what I was talking about.

Answer me this anybody; What socket has lasted 5+ years? Based on the launch date and the date of the last uarch update? AM4 has been the most compatible for the longest period of time. Socket 7 comes to mind but it just wasn't the same. You would probably have to include Super 7 and I'm not all that familiar about chipset compatibilty then.

But, this is off topic and it seems silly to detail specifics about AM4 in a Zen 4 thread. I will not stray off topic anymore.

 

[Schmide]

Answer me this anybody; What socket has lasted 5+ years?

AM3 lasted 7 years, more if you count the overlap with am2. AM2(+) to AM3 and AM3(+) was basically the same with breaking points between ddr2/3 support and bulldozer not supporting ddr2. There were even ddr2/3 AM3 motherboards that extended processor support back to AM2.

 

[Ajay]

Sure, it will live on for years as there isn't much reason for people on Zen 3 to upgrade anytime soon.

Well, no really good reason. I upgraded to Zen3 this year instead of waiting to Zen5. I’m happy with the improvements, for my purposes. Now, if I had $1500 burning a hole in my pocket - I’d upgrade in a heartbeat, just because I’m a computer geek.

 

[Thunder 57]

AM3 lasted 7 years, more if you count the overlap with am2. AM2(+) to AM3 and AM3(+) was basically the same with breaking points between ddr2/3 support and bulldozer not supporting ddr2. There were even ddr2/3 AM3 motherboards that extended processor support back to AM2.

Are you not reading what I am saying? It's not hard. From the introduction of a new socket to the last uarch for it. AM3 came out in 2009 and stopped at Pildedriver in 2012. And AM2/AM3 were not compatible. Different memory and different pinout.

Socket AM3 breaks compatibility with AM2/AM2+ processors due to a subtle change in key placement. The AM3 socket has 941 pin contacts in a different layout[3]

AM2 was great. I have said before how you could go from a single core to a dual core to a quad core. I got one of those two quad cores, the Phenom II 940. On AM2, the L3 ran at a reduced speed. An acceptable comprimise, but not something AM4 had.

Well, no really good reason. I upgraded to Zen3 this year instead of waiting to Zen5. I’m happy with the improvements, for my purposes. Now, if I had $1500 burning a hole in my pocket - I’d upgrade in a heartbeat, just because I’m a computer geek.

Same boat as you. Went from a 2600X to a 5700X.

 

[Schmide]

Are you not reading what I am saying? It's not hard. From the introduction of a new socket to the last uarch for it. AM3 came out in 2009 and stopped at Pildedriver in 2012. And AM2/AM3 were not compatible. Different memory and different pinout.

yeah. And there were frankenboards that bridged the gap between am2+ and am3. (and enough with the read what I'm saying. No one needs that)

https://fudzilla.com/news/automotive/15101-msi-does-a-790gx-combo-board

 

[LightningZ71]

Just realize that, at any time they wanted to, they could release a 5XXX+ line for AM4 based around N6 versions of the existing Zen3 CCDs with zero other changes and boost the MT performance of the AM4 platform by a notable amount, especially the 5900/5950. It wouldn't take a lot of investment either as N6 is supposed to be a "low effort" transition for existing N7 designs.

 

[Ajay]

Just realize that, at any time they wanted to, they could release a 5XXX+ line for AM4 based around N6 versions of the existing Zen3 CCDs with zero other changes and boost the MT performance of the AM4 platform by a notable amount, especially the 5900/5950. It wouldn't take a lot of investment either as N6 is supposed to be a "low effort" transition for existing N7 designs.

But they won’t.

 

[Geddagod]

Just realize that, at any time they wanted to, they could release a 5XXX+ line for AM4 based around N6 versions of the existing Zen3 CCDs with zero other changes and boost the MT performance of the AM4 platform by a notable amount, especially the 5900/5950. It wouldn't take a lot of investment either as N6 is supposed to be a "low effort" transition for existing N7 designs.

AMD doesn't use EUV N7 for zen 3 or zen 2 AFAIK. It would still be relatively easy though, since they did use 6nm for Rembrandt. It's "low effort" but also likely low perf gains to boot, so there's really no point.

 

[moinmoin]

I was clearly talking about socketed AM4 boards.

What are you on about, AM4 is a socketed platform. If it has no sockets, it gets a different name and no longer is AM4.

 

[coercitiv]

Answer me this anybody;

I will not stray off topic anymore.

It's too late for that, you have summoned ANYBODY!

AMD’s first AM4 motherboards still getting firmware updates 6 years and 8 months after launch

A quick investigation into this claim has confirmed that all major motherboard manufacturers have indeed released updates in the past 3-4 months. These updates include the latest AGESA firmware update along with some security enhancements.

Seriously now though, AMD clearly sees AM4 as a platform that can still sell well in the world. Not all markets require modern hardware. Going back to the topic that sparked this conversation, Threadripper is nothing like the mainstream platforms, any analogy is moot and you really didn't need to engage with @igor_kavinski and his "TR is dead to me" poetic take.

AMD learned many lessons with AM4, I hope they put this experience to good use with AM5.

 

[igor_kavinski]

AMD doesn't use EUV N7 for zen 3 or zen 2 AFAIK. It would still be relatively easy though, since they did use 6nm for Rembrandt. It's "low effort" but also likely low perf gains to boot, so there's really no point.

Hey, maybe we get it on AM4's 10 year anniversary

 

[eek2121]

Purely gaming is a terrible use case for TR. It is a workstation CPU that can also game well. I also doubt "it would be bought by a lot of people who don't really need it". Just because you want it on the cheap doesn't mean AMD is that stupid. You also failed to address the fact that no socket has lasted five years.

Agreed, and my use case is workstation + gaming. That being said, I have actually done a ton of hobbyist game engine development, and even some multithreaded stuff. I would love to play around with building something that can take advantage of 32+ cores.

Game developers haven’t really explored this area much. Most developers use existing engines and might extend them.

Things like AI, game logic, world event simulation, world streaming, and even ray tracing all benefit from multiple cores/threads.

Just look at what lumen does with the CPU cores in UE5. Imagine if Cyberpunk 2077 actually simulated the entire game world in real time via multiple threads. Rather than cops magically spawning near you, they would get in their cars and drive from their current locations.

Immersion would explode in a world where high core counts are the norm.

That is one of the neat things about Intel’s approach. Intel is giving us 24 cores on the desktop. However, until consoles see higher core counts, we won’t see much innovation.

AMD doesn't use EUV N7 for zen 3 or zen 2 AFAIK. It would still be relatively easy though, since they did use 6nm for Rembrandt. It's "low effort" but also likely low perf gains to boot, so there's really no point.

A 6nm version of Zen 2 also exists.

 

[Geddagod]

A 6nm version of Zen 2 also exists.

wow that's news to me, thanks. That's wild lol
Really curious about why AMD would bother doing that...

 

[Kepler_L2]

wow that's news to me, thanks. That's wild lol
Really curious about why AMD would bother doing that...

Even a 4nm version of Zen2 exists

 

[rtxtwt]

Dudes from computerbase found that AM5 APUs(or without graphic?) seems to be Phoenix 2

BIOS-Update: AMD Phoenix APU könnte bald im Desktop landen

Ein neues BIOS schürt große Erwartungen: Augenscheinlich bringt das AGESA 1.0.8.0 Unterstützung für die ersten APUs der neuen Plattform.

www.computerbase.de

googlish:

The extracted microcode of the BIOS update does not contain the CPUID 00A70F41 for Phoenix. Instead, the CPUID 00A70F80 is listed, which is assigned to Phoenix 2. On the other hand, the CPUID 00A70F52 suggests a Phoenix refresh. The latter, codenamed Hawk Point, had already been pointed out by the mini-PC manufacturer Minisforum.

 

[igor_kavinski]

Even a 4nm version of Zen2 exists

Codename?

 

[Thunder 57]

yeah. And there were frankenboards that bridged the gap between am2+ and am3. (and enough with the read what I'm saying. No one needs that)

https://fudzilla.com/news/automotive/15101-msi-does-a-790gx-combo-board

You're right, no one needs that. Sorry about that.

What are you on about, AM4 is a socketed platform. If it has no sockets, it gets a different name and no longer is AM4.

Again I was wrong. When I read embedded I immediatly thought BGA boards.

It's too late for that, you have summoned ANYBODY!

AMD’s first AM4 motherboards still getting firmware updates 6 years and 8 months after launch


Seriously now though, AMD clearly sees AM4 as a platform that can still sell well in the world. Not all markets require modern hardware. Going back to the topic that sparked this conversation, Threadripper is nothing like the mainstream platforms, any analogy is moot and you really didn't need to engage with @igor_kavinski and his "TR is dead to me" poetic take.

AMD learned many lessons with AM4, I hope they put this experience to good use with AM5.

AM4 has been amazing and I am surprised new BIOS are coming out so much later. I figured after the 300 series got Zen 3 support they would call it a day.

Unfortunately I am not sure how much AMD learned from prior mistakes. Because AM4 was wildly successful they were able to force LGA onto mobo makers. That's a positive. I think they shot themselves in the foot by forcing the socket to be compatible with AM4 coolers. It's a plus early on I suppose. But I don't see how AMD gets to 24 cores without either using ZenC cores or shrinking the I/O and compute dies. Will be intersting to see what happens.

 

[igor_kavinski]

I have actually done a ton of hobbyist game engine development

Open source or your own? Have a github link?

 

[StefanR5R]

Things like AI, game logic, world event simulation, world streaming, and even ray tracing all benefit from multiple cores/threads. [...] Immersion would explode in a world where high core counts are the norm.

That is one of the neat things about Intel’s approach. Intel is giving us 24 cores on the desktop. However, until consoles see higher core counts, we won’t see much innovation.

Both Intel and AMD offer 32 hardware threads on the desktop = 32 logical CPUs for game engine developers to work with. (For the time being, AMD's hardware threads happen to accomplish more than Intel's within a given power envelope.) However, for simulation of complex worlds, you also would have a good use for lots and lots of memory bandwidth. This brings you into territory of... (no, not Threadripper, rather: ) ... EPYC "X" SKUs (i.e. the ones with stacked cache, which helps with the mere twelve memory channels per socket) and Xeon "Max" SKUs (i.e. the ones with HBM, which helps with the mere eight memory channels per socket).

 

[jpiniero]

Unfortunately I am not sure how much AMD learned from prior mistakes. Because AM4 was wildly successful they were able to force LGA onto mobo makers. That's a positive. I think they shot themselves in the foot by forcing the socket to be compatible with AM4 coolers.

AMD screwed up bad by not supporting DDR4 on AM5.

 

[StefanR5R]

More Zen 4 Threadripper news leaking out. 5.1ghz max boost for the 96 core chip and 5.35ghz for the rest of the stack.

https://videocardz.com/newz/amd-thr...-35-ghz-clocks-350w-tdp-and-ddr5-5200-support

The 32 and 64 core parts look pretty impressive.

[Videocardz cite Wcctftech, FWIW.]

I find it remarkable that this rumor speaks of 128 PCIe lanes implemented by the alleged 96-core SKU, but 64 PCIe lanes by all other alleged SKUs. (Epyc 8004's partially disabled IOD offers 96 PCIe lanes, compared to Epyc 9004's 128 lanes.)

Will AMD pull an Intel and offer processors with different PCIe lane count for one and the same socket and platform? Or might there be two platforms coming up? (Which Intel also had during one of their HEDT/WS generations; IIRC it was the one with the infamous don't-mind-the-chiller-under-the-table-cloth demo.)

 

[Markfw]

[Videocardz cite Wcctftech, FWIW.]

I find it remarkable that this rumor speaks of 128 PCIe lanes implemented by the alleged 96-core SKU, but 64 PCIe lanes by all other alleged SKUs. (Epyc 8004's partially disabled IOD offers 96 PCIe lanes, compared to Epyc 9004's 128 lanes.)

Will AMD pull an Intel and offer processors with different PCIe lane count for one and the same socket and platform? Or might there be two platforms coming up? (Which Intel also had during one of their HEDT/WS generations; IIRC it was the one with the infamous don't-mind-the-chiller-under-the-table-cloth demo.)

From what I hear, threadripper 96 cores is SP5 based on 9654 Genoa, where all the rest will be the smaller SP6 socket, and thats why less IOD and less PCIE lanes.

Crap IMO