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!

 

[BorisTheBlade82]

@Panino Manino
To me it is bloody certain that AMD will have more than just thought about bringing their C cores to the mix right now. The question is if they will do something like @Timorous mentioned already in this generation or if it will be a thing for the next one. Technically there is no reason not to do it right now. Introducing 3D Cache with Zen 3 was much more of a stretch than connecting a Zen4c 16c CCD with the exact same interconnect to the consumer IOD.
From a market point of view there is no immediate pressure - but the lead in MT has decreased significantly.

 

[HurleyBird]

3D Cache helps, but helps more with things that Zen 4 is not far behind Raptor Lake.

You make it sound like there's some abundance of scenarios where Zen 4 is far behind Raptor Lake. On the high end at least, the two products are very, very close. Assuming that there is no massive frequency hit, a 7950X3D should edge the 13900K by a substantially larger margin than the 13900K edges the 7950X.

Where Zen 4c might help is lower in the stack where Zen 4 is kind of pathetic. But you would still need a regular Zen 4 chiplet to not lose terribly in single threading, and the silicon cost of one Zen 4 chiplet and one Zen 4c chiplet isn't going to be too dissimilar from two Zen 4 chiplets. Either way the cost isn't that high and, even without Zen 4c, a less greedy AMD could have not released any products with less than 8 cores, and instead created multiple 12 and 16-core SKUs stratified by clockspeed. Such a theoretical lineup would have much better positioning against Raptor Lake. Unfortunately, despite AMD having a much more competitive product this time around, we're currently stuck on a repeat of Zen 3 vs Alder Lake, where only the top x950X SKU makes any sense.

 

[DrMrLordX]

7950X is winning most MT productivity benchmarks vs 13900k, so the idea of adding more cores to be competitive seems wasteful and silly. A 7950X3D would run the table on nearly everything.

 

[poke01]

, so the idea of adding more cores to be competitive seems wasteful and silly

mm not so. Its quite useful in the mid range i5/i7 area.

 

[poke01]

7950X is winning most MT productivity benchmarks vs 13900k

Well it better be, it's more expensive and on a better node...

 

[DrMrLordX]

mm not so. Its quite useful in the mid range i5/i7 area.

You don't understand. I mean that it would be wasteful to add "small" cores to Zen4.

Well it better be, it's more expensive and on a better node...

Agreed. Intel has inferior designs and process nodes.

 

[uzzi38]

mm not so. Its quite useful in the mid range i5/i7 area.

Think about AMD's core layout here. With the two seperate CCDs, you need to have all 8 big cores on a single one so that you can connect them together by a ring bus. Adding on little cores would have to be through a second CCD, with (let's just say) 16 little cores on it.

That would also mean that the r5 and r7 that would use those little cores would also have to be 2CCD products. Frankly at that point they might as well go with an 8 core r5 and a 12 core r7, would still be cheaper than every product requiring 2 CCDs.

 

[Timorous]

7950X is winning most MT productivity benchmarks vs 13900k, so the idea of adding more cores to be competitive seems wasteful and silly. A 7950X3D would run the table on nearly everything.

A 7950X3D will only benefit games and a few niche workloads and for the rest won't really offer much of an MT uplift and if it requires a clock regression might actually be a tiny bit worse in all core MT workloads.

OTOH a '7970X3D' with an 8c Zen 4 CCD with v-cache and a 16c Zen 4c CCD would absolutely dominate everything, it has the v-cache ccd for workloads where cache is king and it has additional cores for workloads where cores are king. It would need better scheduling or some forced profiles in Ryzen Master to work at its best but that is probably doable. This would make an excellent refresh product IMO.

Think about AMD's core layout here. With the two seperate CCDs, you need to have all 8 big cores on a single one so that you can connect them together by a ring bus. Adding on little cores would have to be through a second CCD, with (let's just say) 16 little cores on it.

That would also mean that the r5 and r7 that would use those little cores would also have to be 2CCD products. Frankly at that point they might as well go with an 8 core r5 and a 12 core r7, would still be cheaper than every product requiring 2 CCDs.

I don't think that will even work really. If the 7600X was the 8c 7700X it would still lose in apps to the 13600K.

It might be a bit late for the Zen 4 range but with Zen 5 I think all products will be dual CCD. One will be the Zen 5 CCD and the other the Zen 5c CCD.

Something like this seems like a sensible stack.

Main CCD core count	compact CCD core count	total	Tier	Price??
8 + vcache	16 (+ vcache just to cement halo status maybe?)	24	R9	$800
8 + vcache	12	20	R9	$600
8 + vcache	8	16	R7	$500
6	8 (maybe 6 here if they want more delta to the R7)	14	R5	$350
6	4	10	R5	$280
6	0	6	R3	$230

So if we applied this to the current tier the 7600X would be an R3 probably a 7300X. You could have an R5 7500X as the 6+4 config and an R5 7600X as the 6+8 config. 7800X would be the 8+8 config, 7900X would be the 8+12 config and the 7950X would be the 8+16 config. This would mean no 7700X but perhaps that could be the same config as the 7800X but without the v-cache.

What is clear is that this lineup would absolutely stomp RPL every conceivable way. The 7800X if configured as above would be clearly the faster gaming CPU vs the 13900K and it would probably trade blows with the 13900K in apps while being a small amount cheaper. The 7900X would be there to take the undisputed crown and the 7950X would be there as a halo product that Intel would not be able to compete with at all.

For Zen 4 though I don't think any of this will happen. I could maybe see a new top SKU once Zen 4c is in mass production but other than that I think AMD will just adjust pricing. Based on apps and gaming if AMD fix the dual CCD gaming issues somehow (Ryzen Master profiles to force threads to the good CCD for those games impacted perhaps?) then the 7900X needs to match the 13700K in price and then the competition is pretty even all around except AMD have the perf/W advantage. The 7700X probably needs to match the 13600X in price and the 7600X needs to be closer to $250 or maybe less. Even with the 7700X at 13600K prices the decision will be gaming vs apps because while the 7700X would win by a small bit in gaming the 13600K is very impressive in MT apps so there is a viable reason to go either way depending on use case. Again though AMD have the perf/W advantage.

Then there is the 7800X3D that is likely coming to take the gaming crown. AMD could probably price this between the 13700K and 13600K given it would handily win in gaming but handily loses in apps with 7950X3D (and maybe 7900X3D if that actually exists) being the can do it all halo part although as I said above a 7970X3D with just 1 v-cache CCD and 1 Zen 4c CCD would be even better as long as threads get allocated correctly.

 

[poke01]

The r5 needs to move to 12 cores. r7 to 16 and r9 to 24 cores.

Agreed. Intel has inferior designs and process nodes.

Yep, they will be like that till Arrow Lake.

 

[DrMrLordX]

@Timorous

I can't quote you since some stupid table keeps imposing itself on the reply; regardless, it's not likely that v-cache will come with a clockspeed regression this time around. The 7950X already is able to sustain clocks with a mild undervolt as it is, so AMD has some room to play with binning etc.

 

[JustViewing]

Still I think AMD missed the trick by not introducing 24 core Zen4 CPU. They can charge whatever they want to this product, since it will have very fast low thread performance and unbeatable multi thread performance. I believe there is enough space for 3CCDs in the Zen4 socket layout. If this is not true, then AMD has scarified performance in the name of cooler compatibility. Which is a very sad compromise in my opinion. Which also means that AM5 won't have much longevity or it would be hamstrung for future iterations. IOD should be able support 3 CCD's since it a quater of Epyc's IOD.
Another bad trend I see(at least in my perspective) is reducing number of PCIe slots. I hardly seen any new Mother Boards with over 3 PCIe slots. I think MB manufactures should stack NVME's to save space or some other creative ways to support mutiple NVME's. Even my lowly B350 has 6 slots, out of it I am already using 3 PCIe slots.

 

[Timorous]

@Timorous

I can't quote you since some stupid table keeps imposing itself on the reply; regardless, it's not likely that v-cache will come with a clockspeed regression this time around. The 7950X already is able to sustain clocks with a mild undervolt as it is, so AMD has some room to play with binning etc.

Even if clocks are the same a 7950X3D won't offer much of an MT benefit outside of niche workloads.

The good news is AMD have lots of options even if not all of them are available immediately. They have ways to improve MT performance through a 3rd CCD or a 16c Zen 4c CCD. They have 3d cache to improve gaming performance and niche workload performance and the CCDs are also tiny so it is not exactly expensive to manufacture so have pricing flexibility as well.

Like I say I don't think we will see Zen 4c in desktop unless AMD decide to release a 7970X3D halo part. I could see this from a software proving pov as well because if AMD decide to make Zen 5 multi CCD across most of the range then they need to work on scheduler / profiles anyway so having a 7970X3D in the wild to help that along ready for Zen 5 might be a smart play. If AMD want to be cute they can name the 3D cache version of RDNA3 the 7970XT3D as well just for extra confusion and naming synergy.

 

[Timmah!]

The r5 needs to move to 12 cores. r7 to 16 and r9 to 24 cores.

Agreed. Gimme that 24C R9

 

[igor_kavinski]

Is there a possibility to 3D stack the IO die below the compute dies, so we can have three compute dies on top?

 

[eek2121]

@Panino Manino
To me it is bloody certain that AMD will have more than just thought about bringing their C cores to the mix right now. The question is if they will do something like @Timorous mentioned already in this generation or if it will be a thing for the next one. Technically there is no reason not to do it right now. Introducing 3D Cache with Zen 3 was much more of a stretch than connecting a Zen4c 16c CCD with the exact same interconnect to the consumer IOD.
From a market point of view there is no immediate pressure - but the lead in MT has decreased significantly.

AMD has already said Zen4c is only for cloud. It is not coming to desktop.

 

[NostaSeronx]

I hope you guys don't curse this world with a Mongoose-Zen M6/ZenM thing. Where it is 16 1T cores with a SoIC Logic-on-Logic L3 thingamajig. Stacked 4-Hi to get 64 cores, just for AMD to say they have 2x the little cores. What are they are going to even call it? Zen4-XE3D?!

 

[igor_kavinski]

I hope you guys don't curse this world with a Mongoose-Zen M6/ZenM thing.

Why would it be a curse?

 

[Timorous]

I hope you guys don't curse this world with a Mongoose-Zen M6/ZenM thing. Where it is 16 1T cores with a SoIC Logic-on-Logic L3 thingamajig. Stacked 4-Hi to get 64 cores, just for AMD to say they have 2x the little cores. What are they are going to even call it? Zen4-XE3D?!

This just reads like a Star Trek TNG script.

 

[Zepp]

This just reads like a Star Trek TNG script.

"O'Brien is the ZEN4-XE3D online yet?"

"No sir, I had to re-calibrated the phase modulator, but the power coupler was emitting a high phase variance."

"Try reversing the polarity?"

"Aye Sir"

 

[nicalandia]

So it Begins...!!! Total Domination

https://www.amd.com/en/events/epyc

 

[Panino Manino]

AMD has already said Zen4c is only for cloud. It is not coming to desktop.

And they're wrong!

 

[DrMrLordX]

Even if clocks are the same a 7950X3D won't offer much of an MT benefit outside of niche workloads.

It doesn't need to.

 

[Saylick]

So it Begins...!!! Total Domination

View attachment 69774

https://www.amd.com/en/events/epyc

So we'll finally see Genoa and likely Bergamo, but what's the odds we see MI300?

 

[nicalandia]

So we'll finally see Genoa and likely Bergamo, but what's the odds we see MI300?

Not Likely, but If We do..!!! Could MI300 Be used as a Super Charged UHDT? I mean it's a Super APU.

 

[HurleyBird]

Frankly at that point they might as well go with an 8 core r5 and a 12 core r7, would still be cheaper than every product requiring 2 CCDs.

It's baffling AMD didn't do this in the first place. They knew they were going to be extremely uncompetitive with low end RL, they knew demand had fallen off a cliff well before the Zen 4 launch, and they're still cost competitive even with 2 CCDs, but they went ahead and launched the 7600X and 7700X anyway. Who do they think they are, Nvidia?