Question Zen 6 Speculation Thread

[adroc_thurston]

A full cache Zen 5 exists today with 128 cores. It only makes sense for there to be a Full cache Zen 6 with at least 128 cores.

There are loads where full cache versions of Zen will perform better than Zen c versions will. This is the reason that AMD had Venice full in the first place. Dense serves different computing needs.

Should AMD only offer 96c for Zen 6 full cache version, that would indeed be a step backwards from a full cache Zen5 with 128 cores.

they're all 4M L3$ per core now.

There are loads where full cache versions of Zen will perform better than Zen c versions will.

dense just means lower fmax, L3$ config can be whatever.

There will be (IMO) a 128c Zen 6 full cache version.

It does not exist.
Classic tops out a 96 cores. that's it, pack it up.

Time will tell.

There is no time, platforms are frozen, parts have taped out awhile ago.

 

[Tigerick]

A full cache Zen 5 exists today with 128 cores. It only makes sense for there to be a Full cache Zen 6 with at least 128 cores.

There are loads where full cache versions of Zen will perform better than Zen c versions will. This is the reason that AMD had Venice full in the first place. Dense serves different computing needs.

Should AMD only offer 96c for Zen 6 full cache version, that would indeed be a step backwards from a full cache Zen5 with 128 cores.

Saying that Zen 6c may be as potent as Zen 5 full doesn't change things either. Zen 6 full will exist and there will be customers with loads that would do better on more cores of Zen 6 full than 96. In fact, it may well be that the 128c Turin might best the 96c Zen 6 full in quite a number of benchmarks and applications. Hard to see that as anything but a step back ..... which is also the reason I believe AMD wont be doing that. There will be (IMO) a 128c Zen 6 full cache version.

Time will tell.

You do have a point here. Guess what? There was leaks about Zen6 and Zen6c core count and power consumption in the past; I can't seem to locate it. But I did remember there are 8 x Zen6 (96-core) with 400W TDP and 4 x Zen6c (128-core) with 350W TDP. Both should be available for SP8 platform.

 

[basix]

I do also not see it as a dramatic change, if non-dense cores stop at 96C. Two reasons for that:

• One popular SKUs is e.g. the 64C 9575F. Zen 6 and N2(P) might allow high frequency variant for the 96C SKU
  • Many applications like single threaded performance. Or the software licensing stuff kills core count benefits
  • With "only" 96C you can increase the clock rates. The 128C Zen 5 model clocks only at 4.1 GHz whereas the 96C model clocks at 4.5 GHz
  • Would be funny: If AMD releases Zen 6 "F" variants with 6.0 GHz
• N2(P), NanoFlex and the same L3$ size could bring Zen 6c closer to the non-dense variant. I expect ~4.5 GHz boost clockrate for Zen 6c and ~5 GHz or a bit above for the non-dense version (maybe 5.3 GHz?)
  • 5 GHz vs. 4.5 GHz would be a similar difference like 4.5 GHz to 4.1 GHz today on Zen 5. We would have the same situation as today regarding clockrate difference of 96C and 128C SKU. And you get the full L3$ per core also on Zen 6c. From outside it would look the same, just that you can use Zen 6c cores on the 128C model.

Another point to consider is Zen 7. Rumours say the core count will get increased to 15...16 (hmm, a 3x5 or better 4x4 core layout would make sense then 4x8+ for the bigger CCD ) and therefore we get back to 120...128C again.

There was leaks about Zen6 and Zen6c core count and power consumption in the past; I can't seem to locate it. But I did remember there are 8 x Zen6 (96-core) with 400W TDP and 4 x Zen6c (128-core) with 350W TDP

350W to 400W sounds about right. We have 400W and 400W for the two different Zen 5/5c 96C/128C SKUs. But as above written, I expect clockrates to increase for the 96C model. That could explain the difference. Or N2 delivers some magic with its dense libraries or 350W is simply much heat dissipation for just 4x 32C chiplets. The 256C model with 700W would then deliver the exact same clockrates as the 128C model. Does not sound too bad for me.

 

[adroc_thurston]

bravo
r
a
v
o

finally someone with a brain figured it out.

 

[marees]

I like how he states Zen 6 is just an evolution of Zen 5. No hype trains in sight.

Regarding the those memory controllers, It's known upcoming server CPUs support both DDR5 and MR-DIMM.

How about HB-DIMM ?

AMD patent details HB-DIMM RAM implementation that doubles current DDR5 bandwidth for consumer PCs

Team Red's HB-DIMM standard is designed to effectively double the current DDR5 bandwidth by reaching 12.8 Gbps transfer speeds, but without radically changing the design in order to ensure compatibility with current consumer and server-grade systems.

www.notebookcheck.net

 

[OneEng2]

• One popular SKUs is e.g. the 64C 9575F. Zen 6 and N2(P) might allow high frequency variant for the 96C SKU
  • Many applications like single threaded performance. Or the software licensing stuff kills core count benefits
  • With "only" 96C you can increase the clock rates. The 128C Zen 5 model clocks only at 4.1 GHz whereas the 96C model clocks at 4.5 GHz
  • Would be funny: If AMD releases Zen 6 "F" variants with 6.0 GHz
• N2(P), NanoFlex and the same L3$ size could bring Zen 6c closer to the non-dense variant. I expect ~4.5 GHz boost clockrate for Zen 6c and ~5 GHz or a bit above for the non-dense version (maybe 5.3 GHz?)
  • 5 GHz vs. 4.5 GHz would be a similar difference like 4.5 GHz to 4.1 GHz today on Zen 5. We would have the same situation as today regarding clockrate difference of 96C and 128C SKU. And you get the full L3$ per core also on Zen 6c. From outside it would look the same, just that you can use Zen 6c cores on the 128C model.

Thanks for the detailed response.

As you point out, there are workloads that charge by the core and applications (like databases) that thrive on strong single core performance.

While Zen 6 will most certainly offer improvements over Zen 5, it seems unlikely that the clock speed and IPC improvements will overcome the 33% difference in core count.... thus making this a "step back" for AMD.

Another point to consider is Zen 7. Rumours say the core count will get increased to 15...16 (hmm, a 3x5 or better 4x4 core layout would make sense then 4x8+ for the bigger CCD ) and therefore we get back to 120...128C again.

Oh, I totally agree that should AMD walk backwards with Venice (from Turin) on full core counts, that they will return those cores in Zen 7. I am just curious how a core count reduction of full Zen 6 is not considered a step backwards?

Do we expect more than a 33% per core performance boost from Zen 5 to Zen 6?

 

[techjunkie123]

they're all 4M L3$ per core now.

dense just means lower fmax, L3$ config can be whatever.

So dense has the same L3/core as classic this time, instead of half?

I wonder what the difference is size is between the classic and dense then.

 

[CouncilorIrissa]

So dense has the same L3/core as classic this time, instead of half?

I wonder what the difference is size is between the classic and dense then.

The core area difference has never taken L3 into consideration to begin with. So it's due to lower fmax (means you can use denser libraries/2-1 or 2-2 fin cells in the case of Finflex), same as it's always been.

 

[StefanR5R]

@OneEng2, to put something which has been said by others a few times already into yet another wording:
a) There is something which EPYC 9755 offers which EPYC 9745 does not,
b) yet there is nothing which a 128c Zen6-classic SoC would offer which a 128c Zen6-dense SoC will not.

 

[eek2121]

Thanks for the detailed response.

As you point out, there are workloads that charge by the core and applications (like databases) that thrive on strong single core performance.

While Zen 6 will most certainly offer improvements over Zen 5, it seems unlikely that the clock speed and IPC improvements will overcome the 33% difference in core count.... thus making this a "step back" for AMD.

Oh, I totally agree that should AMD walk backwards with Venice (from Turin) on full core counts, that they will return those cores in Zen 7. I am just curious how a core count reduction of full Zen 6 is not considered a step backwards?

Do we expect more than a 33% per core performance boost from Zen 5 to Zen 6?

They aren’t walking backwards with anything. Dense will replace the previous SKUs and improve performance.

 

[jpiniero]

b) yet there is nothing which a 128c Zen6-classic SoC would offer which a 128c Zen6-dense SoC will not.

Then why offer Classic?

 

[511]

Then why offer Classic?

For GPUs or stuff that favours fewer but higher performing cores.

 

[OneEng2]

Then why offer Classic?

... and thus my query .

Either:

A) Classic offers nothing over Dense in Zen 6 - Then why offer it?

OR

B) Zen 6 Classic having 33% fewer cores than Zen 5 Classic is a step backwards.

 

[inquiss]

... and thus my query .

Either:

A) Classic offers nothing over Dense in Zen 6 - Then why offer it?

OR

B) Zen 6 Classic having 33% fewer cores than Zen 5 Classic is a step backwards.

Or c) cloud providers want lots of dense cores so they get that and others want more frequency optimised cores so they get that. It's c) everyone is telling you it's c). For the record, it's c)

 

[StefanR5R]

Then why offer Classic?

Are you referring to ≥96c/socket? Or is your question more general?

A) Classic offers nothing over Dense in Zen 6 - Then why offer it?

You removed important context.

 

[jpiniero]

Are you referring to ≥96c/socket? Or is your question more general?

The whole thing is dumb and I assume the "leakers" have it wrong. Yet again.

 

[basix]

B) Zen 6 Classic having 33% fewer cores than Zen 5 Classic is a step backwards.

You are drawing the wrong picture. Yes, you have less "classic" cores. But you do not have a performance regression in any way:

• Disclaimer: Speculated frequencies to bring my point forward

	Core count	Boost clock	Base clock	L3$ per core
Z5 classic	96	4.5 GHz	2.6 GHz	4 MByte
Z5 classic	128	4.1 GHz	2.7 GHz	4 MByte
Z5 dense	128	3.7 GHz	2.4 GHz	2 MByte
Z6 classic	96	5.2 GHz	3.5 GHz	4 MByte
Z6 dense	128	4.5 GHz	3.0 GHz	4 MByte

As you can see, clock rates go similarly up for all core counts. And because Zen 6c does feature the full cache capacity, you can directly replace the 128C model of Z5 classic with Zen 6c.

Then why offer Classic?

5+ GHz SKUs won't be possible without Classic. And there are lower coure count models (48 core etc.) which could clock higher than just 4.5 GHz of what Zen 6c might bring to the table.
Using one or more 32C chiplets would also not be economical for many lower core count SKUs. Use one ~140mm2 Die or 2x ~60...70mm2 makes a difference in cost and yield (e.g. 24C / 48C models)

And last but not least:
Consumers and gam0rz! They like high frequencies

 

[lucasworais]

Also, if there is low or no demand for a certain SKU, you kill it. Is that simple.

Maybe a 128c Zen6 would offer more than 128 Zen6c, but if no one wants it why even bother?

 

[basix]

Zen 6 classic could only deliver one single thing: Higher frequencies.

I think the market for that at 128C gets pretty slim. The only really cool and differentiatable SKU would be an 128C F-SKU. But the 96C F-SKU will probably already draw 600W (the 64C 9575F sits at 400W).
So the only really interesting SKU will not be technically viable within the SP7/SP8 socket specifications at probably max. 600W.

 

[adroc_thurston]

The only really cool and differentiatable SKU would be an 128C F-SKU

Not fittable in SP8 which is 350W skt power.

 

[basix]

SP7 then

 

[adroc_thurston]

SP7 then

no classic SP7 exists.

 

[OneEng2]

You are drawing the wrong picture. Yes, you have less "classic" cores. But you do not have a performance regression in any way:

• Disclaimer: Speculated frequencies to bring my point forward

Core count	Boost clock	Base clock	L3$ per core
Z5 classic	96	4.5 GHz	2.6 GHz	4 MByte
Z5 classic	128	4.1 GHz	2.7 GHz	4 MByte
Z5 dense	128	3.7 GHz	2.4 GHz	2 MByte
Z6 classic	96	5.2 GHz	3.5 GHz	4 MByte
Z6 dense	128	4.5 GHz	3.0 GHz	4 MByte

As you can see, clock rates go similarly up for all core counts. And because Zen 6c does feature the full cache capacity, you can directly replace the 128C model of Z5 classic with Zen 6c.

Thanks again for a very thoughtful response.

Let's take Oracle as an example. The licensing cost for on-site Oracle is per core. This licensing cost is far more than the cost of the hardware and therefore pushes companies to the highest power per-core processors out there.

Fanning out more cores cost linearly more money EVERY YEAR and dwarfs the hardware cost.

It is in these situations that a 128c Zen 6 Classic would make sense over a 128c Zen 6c.

Also, if there is low or no demand for a certain SKU, you kill it. Is that simple.

Maybe a 128c Zen6 would offer more than 128 Zen6c, but if no one wants it why even bother?

So, just to be clear, most people here are not arguing that a 96c Venice Zen 6 Classic is better than a 128c Turin Classic, but rather that there is no need in the market for this part at all.

Does anyone have some market research to prove this theory?

I am looking at this from the landscape of competition likely to exist in 2025-2027.

Intel will be introducing Granite Ridge with 192 Panther Cove cores. These will not have SMT and therefore, I am dubious about their ability to compete 1 on 1 with a Zen 6 Classic, so perhaps AMD could get away with neglecting this market for a couple of years.

Note: This argument is quite different from "no one wants it", or "there is no application for it". This argument is that "People might prefer it, but since AMD will still offer the best x86 solution over Intel, people will still buy it because it will be the best solution available at that time".

 

[adroc_thurston]

It is in these situations that a 128c Zen 6 Classic would make sense over a 128c Zen 6c.

you do not want either of those for per-C licensing.
you want a 32 or 48c fmax part.

 

[Markfw]

Note: This argument is quite different from "no one wants it", or "there is no application for it". This argument is that "People might prefer it, but since AMD will still offer the best x86 solution over Intel, people will still buy it because it will be the best solution available at that time".

I am not sure the total market for 128 core Turin, or Venice, but my 9755 is not alone in several applications that it uses and uses 128 core fat (NO SMT) or 256 thin (with SMT), and I just don't know how many do, but it certainly is not zero.

@adroc_thurston can you add your opinion on how big this market is ?