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!

 

[Joe NYC]

Maybe AMD is just winding down Zen 3 chiplet production and it just makes more sense for them to have Zen 4 chiplets going forward for both AM4 and AM5.

Zen 3 chiplet production will keep on going full speed for Milan, Milan-X, in addition to some desktop production for upgraders.

 

[CakeMonster]

Are any knowledgeable people willing to make a bet which models will get the X3D versions with V-Cache?

I've seen speculations and rumors all over the place, but none have been able to make an argument for why it would be those, so I kind of suspect its rehashing rumors or making stuff up. Only the 8c CCX parts would kind of make sense if that's what is produced in the first place, but some have said the 7800X3D and 7900X3D for example. I could see how AMD might want a secondary lineup just with most or all of them too just to take people's money, or a flagship mythical 7950X3D with similar reasoning. I don't know if we have enough info about capacity for producing them and relative cost either.

 

[Zucker2k]

Are any knowledgeable people willing to make a bet which models will get the X3D versions with V-Cache?
I've seen speculations and rumors all over the place, but none have been able to make an argument for why it would be those, so I kind of suspect its rehashing rumors or making stuff up. Only the 8c CCX parts would kind of make sense if that's what is produced in the first place, but some have said the 7800X3D and 7900X3D for example. I could see how AMD might want a secondary lineup just with most or all of them too just to take people's money, or a flagship mythical 7950X3D with similar reasoning. I don't know if we have enough info about capacity for producing them and relative cost either.

The 8core is more than likely, based on the 5xxxX3D. The 12core 5000 series was what AMD touted as the gaming chip, plus the first X3D demo was done with a prototype 12core X3D chip, so based on the above, I won't be surprised to see the 8c and 12c chips getting the X3D treatment.

There could be a cannibalization reason why the 16c isn't getting that treatment because it may eat into the Genoa 16c territory. Who knows. Whatever the case, I believe AMD would release the necessary chips to take the fight to Intel, including the 16c X3D.

 

[Timmah!]

The 8core is more than likely, based on the 5xxxX3D. The 12core 5000 series was what AMD touted as the gaming chip, plus the first X3D demo was done with a prototype 12core X3D chip, so based on the above, I won't be surprised to see the 8c and 12c chips getting the X3D treatment.

There could be a cannibalization reason why the 16c isn't getting that treatment because it may eat into the Genoa 16c territory. Who knows. Whatever the case, I believe AMD would release the necessary chips to take the fight to Intel, including the 16c X3D.

They better be releasing that 16C with v-cache. Its already enough there is no 24C, but not even making that 16C the best it could be, that would be travesty.

 

[Thibsie]

They better be releasing that 16C with v-cache. Its already enough there is no 24C, but not even making that 16C the best it could be, that would be travesty.

Without knowing how Zen4 fares against Intel, I find this kind of comments completely ridiculous.

 

[Abwx]

They better be releasing that 16C with v-cache. Its already enough there is no 24C, but not even making that 16C the best it could be, that would be travesty.

If they want to charge 700-800$ for a 7950X then there s no price bracket that would be available for a 24C part.

 

[Timmah!]

Without knowing how Zen4 fares against Intel, I find this kind of comments completely ridiculous.

What are you talking about. I think its fairly reasonable to assume that v-cache CPUs will be overall superior to vanilla ones, bar significant clock penalty, so if i am looking at 16C, i want one with v-cache. Intel has absolutely nothing to do with that. If i was looking to buy Intel CPU, i would get one and not care about nuances in the AMDs line-up.

If i am buying new rig, its not in order to brag about how better it is than your Intel/AMD/Nvidia/Apple/whatever one, but to have better hardware than my current one. And by better i mean as much as possible, at the price i am willing to spend.

 

[Timmah!]

If they want to charge 700-800$ for a 7950X then there s no price bracket that would be available for a 24C part.

Oh yeah, because they have issues to invent new "price brackets".

Remind me, what "client" CPU was sold at 799 before 16C Ryzens became a thing?

 

[Abwx]

Oh yeah, because they have issues to invent new "price brackets".

Remind me, what "client" CPU was sold at 799 before 16C Ryzens became a thing?

799 or so is the highest you could get in significant amount, with 50% more cores they would have to price it at 1300$ minimaly, that s outside of the enthusiast budget, 799 is already quite a hefty pricing for a consumer dedicated product.

 

[Timmah!]

799 or so is the highest you could get in significant amount, with 50% more cores they would have to price it at 1300$ minimaly, that s outside of the enthusiast budget, 799 is already quite a hefty pricing for a consumer dedicated product.

They clearly trying to build an impression that the 670 "EXTREME" is sort of new "HEDT". There was plenty of "enthusiasts" who own or owned 1000+ X99/X299/X399 CPUs in the past, willing to pay more than 799, but would not go as far as 2500 for cheapest TR. There is massive hole between 800 and 2500 pricepoints now and there would be plenty of people willing to pay for product in that segment, be it for work, gaming, messing around with OCing or simply privilege of owning cutting edge HW/bragging rights.

Intel 6700K, which was a client flagship before first Ryzens dropped, was 350 USD. 6-core i7 8700k was later 359, 8-core 9900k/Ryzen 1800x were 499.
16-core 3950x was then 799. If you say that 24C would be 1300, then the jump would be as big as jump from 1800x to 3950x, + cca 60 percent. In other words it happened before, and mark my word, when 24C eventually drops, maybe with Zen5 family, it will happen again.
They increase the prices, invent the new price brackets and shift the segmentation in their favor all the time and they absolutely dont care whether their halo product its outside of anyones budget.

 

[HurleyBird]

There could be a cannibalization reason why the 16c isn't getting that treatment because it may eat into the Genoa 16c territory. Who knows. Whatever the case, I believe AMD would release the necessary chips to take the fight to Intel, including the 16c X3D.

With 12 compute dies, if Genoa works like past Epyc generations, cores will be enabled in increments of 12, so no 16 core SKU.

 

[moinmoin]

With 12 compute dies, if Genoa works like past Epyc generations, cores will be enabled in increments of 12, so no 16 core SKU.

Fewer compute dies will still be possible. Though the lower core/fewer compute dies amount likely will be covered by the lower cost Siena platform instead later.

 

[Exist50]

They clearly trying to build an impression that the 670 "EXTREME" is sort of new "HEDT". There was plenty of "enthusiasts" who own or owned 1000+ X99/X299/X399 CPUs in the past, willing to pay more than 799, but would not go as far as 2500 for cheapest TR. There is massive hole between 800 and 2500 pricepoints now and there would be plenty of people willing to pay for product in that segment, be it for work, gaming, messing around with OCing or simply privilege of owning cutting edge HW/bragging rights.

Intel 6700K, which was a client flagship before first Ryzens dropped, was 350 USD. 6-core i7 8700k was later 359, 8-core 9900k/Ryzen 1800x were 499.
16-core 3950x was then 799. If you say that 24C would be 1300, then the jump would be as big as jump from 1800x to 3950x, + cca 60 percent. In other words it happened before, and mark my word, when 24C eventually drops, maybe with Zen5 family, it will happen again.
They increase the prices, invent the new price brackets and shift the segmentation in their favor all the time and they absolutely dont care whether their halo product its outside of anyones budget.

I think the trouble with the HEDT comparison is that the mainstream platforms are short on memory bandwidth/capacity and IO by comparison. And ultimately, I wonder what the trend here is. MT CPU performance is growing faster than memory bandwidth. I guess we can keep throwing more cache at the problem, but that'll get expensive fast.

 

[tomatosummit]

799 or so is the highest you could get in significant amount, with 50% more cores they would have to price it at 1300$ minimaly, that s outside of the enthusiast budget, 799 is already quite a hefty pricing for a consumer dedicated product.

A non issue.
The 3950/5950 supplanted the 1950/2950 threadrippers already so the 24 core thread ripper isn't protected.
What a 24 core zen4 cpu would do is increase production cost of all IO die based cpu for the third ifop link and compete primarily against their own 16 core cpus.
I want the 24core as well but the reason it doesn't exist certainly isn't the expensive market position.

 

[MadRat]

If the 12 core is derived from a pair of 8 cores, would you want 6+6 or 4+8 for best single thread performance? If core0 resides on chiplet0 then only having three cores to share on the chiplet0, with core1 through core9 on chiplet1, you have the least used cores sharing chiplet0 with core0. Bonus if the deactivated core caches could be used to supplement core0.

 

[HurleyBird]

Fewer compute dies will still be possible. Though the lower core/fewer compute dies amount likely will be covered by the lower cost Siena platform instead later.

Possible sure, but iirc AMD has only ever done that with TR, not Epyc.

 

[SteinFG]

Possible sure, but iirc AMD has only ever done that with TR, not Epyc.

Nah, they do it on all platforms.

AMD EPYC 9664 could be the flagship 96-core Genoa CPU with 400W TDP - VideoCardz.com

AMD Genoa flagship SKU reportedly named EPYC 9664 Leaker YuuKi_AnS, better known for various pre-launch reviews of unreleased hardware, claims that the flagship of the upcoming AMD Genoa data-center CPU series might be called EPYC 9664. AMD EPYC Genoa CPU The leaker already revealed the list of...

videocardz.com

If leaked lineup is right, only 96-core Genoa has all 12 CCDs, 32-64 core genoa parts have 8 CCDs, and 16-24 core ones have 4 CCDs
Also, Epyc 7002 even had skus with just 2 CCDs, like this

AMD EPYC 7002 Rome CPUs with Half Memory Bandwidth

We delve into why some AMD EPYC 7002 CPUs have half the memory bandwidth of others as 4-channel optimized SKUs and why that is fine for their target market

www.servethehome.com

 

[gruffi]

Don't assume that just because Alder Lake benefits greatly from faster DDR5 that Zen 4 will as well. Alder Lake has only 30mb of L3 cache to share with 16 cores (and the L3 cache isn't very fast) which is why it benefits so heavily from fast DDR5 memory, likely due to all the cache misses in games.

Actually there is not much capacity difference between the L3 cache of one Zen 3 CCD and Alder Lake S, 32 vs 30 MB. Zen 3 also can greatly benefit from faster RAM. https://github.com/xxEzri/Vermeer/blob/main/Guide.md

Large caches reduce the impact of faster memory.

Sure. But that's mainly the strength of the Zen 3/4 V-Cache SKUs, not the normal SKUs.

 

[Kaluan]

Sure. But that's mainly the strength of the Zen 3/4 V-Cache SKUs, not the normal SKUs.

Well, I could see future 24/32 core Zen5/Zen5+ have V-Cache or other 3D stacking as a baseline feature TBH, provided 1. AMD masters this tech enough and 2. It becomes 'cheap' enough to not affect BOM/packaging costs too much anymore.

Edit: Also DDR5 is far from being fully tapped out, with JEDEC specs going up to 7200MTs (vs today's native 4800-5200). So there's also that.

 

[Hans Gruber]

Well, I could see future 24/32 core Zen5/Zen5+ have V-Cache or other 3D stacking as a baseline feature TBH, provided 1. AMD masters this tech enough and 2. It becomes 'cheap' enough to not affect BOM/packaging costs too much anymore.

Edit: Also DDR5 is far from being fully tapped out, with JEDEC specs going up to 7200MTs (vs today's native 4800-5200). So there's also that.

So we learned from B350/X370 that the limitation of memory bandwidth was a bios issue and not an infinity fabric issue. 3800mhz is the 1:1 maximum speed for Zen 2/3.

We need to know what the maximum DDR5 memory speed will be without crazy dividers introducing massive latency penalties. It looks as if 6400mhz DDR5 is what the starting point should be for Zen 4 builds. I expect AMD will scale much better than Intel with fast memory on Zen 4. I also expect to see 8000mhz DDR5 kits at some point in 2023.

 

[LightningZ71]

The most logical reason that there was never a 5950x3d is socket ppwer limitations. We already see significant limitations on click speeds on the 5800x3d when it is run at spec. The 5950x3d would likely never break 4Ghz at spec and would suffer greatly against other chips.

The 7000 series has a greater socket power spec and a far more efficient ccd process and IOD process. It should be able to clock well enough to make it a viable product. In addition, the serdes link for Zen3 should technically be significantly slower than the one on Zen4, if for no other reason than to allow Zen4 to make full use of the higher speed DDR5 memory. That higher serdes link speed also means that core-core communications should be more efficient across those links and better allow for cache coherency between the CCDs when equipped with a large VCACHE.

Its my belief that the higher socket power isn't specifically forvtge 7950x, but also to support the 7950x3d at higher speeds.

 

[Joe NYC]

Oh yeah, because they have issues to invent new "price brackets".

Remind me, what "client" CPU was sold at 799 before 16C Ryzens became a thing?

 

[Gideon]

Are any knowledgeable people willing to make a bet which models will get the X3D versions with V-Cache?

I've seen speculations and rumors all over the place, but none have been able to make an argument for why it would be those, so I kind of suspect its rehashing rumors or making stuff up. Only the 8c CCX parts would kind of make sense if that's what is produced in the first place, but some have said the 7800X3D and 7900X3D for example. I could see how AMD might want a secondary lineup just with most or all of them too just to take people's money, or a flagship mythical 7950X3D with similar reasoning. I don't know if we have enough info about capacity for producing them and relative cost either.

I'm personally rooting for both the single-chiplet and dual-chiplet models getting V-cache.

I would really like to see a specialized cache die overlapping BOTH chiplets, allowing for insanely sized unified caches (something similar to what they had in their GPU patents). This is probably not gonna happen though. While server products could justify the expenses, just adding more layers of V-cache on top of the existing one would be a far easier approach for them (unfortunatly this is also less usable in ultra-high frequency gaming CPUs).

 

[Hans Gruber]

Is there any chance AMD will release a 24core or 32core Zen 4 CPU during the product cycle? What is the maximum theoretical core count for a Zen 4 CPU's on the consumer grade motherboard/chipset?

 

[RnR_au]

Is there any chance AMD will release a 24core or 32core Zen 4 CPU during the product cycle? What is the maximum theoretical core count for a Zen 4 CPU's on the consumer grade motherboard/chipset?

Is there space for more than 2 8 core dies? Along with the IO die?