Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[BorisTheBlade82]

Yikes, looks like the rumor of Zen 5 + Zen 4D could be true after all
View attachment 78615
View attachment 78612
Reading Family ID is not sufficient anymore, now there are core level differences and the CPU Topology has to be read as well.

Amazing, were did you find that?
Slowly but surely all my wild speculations come true.
Next stop: IFoP successor 😎

 

[Timmah!]

Same thought, I only want the Zen 5 cores, not some puny outdated Zen 4c inebench cores.


I would guess one 8C Zen 5 Chiplet + 16C Zen 4 Chiplet for 24C.
So going by PHX2 topology, Zen 4c would have half the L3 per Core but for 16C in one CCX the amount of L3 would still be same? That could make L3 available to a core to the SW pretty much same even though L3 per core is half. But overall around same size as Zen 4 CCD but 2x cores, if going with PHX density of 25 MTr@178mm2
View attachment 78613

Meh. Give me 24 Zen5 cores not these hybrids. If i wanted that, i could get Intel CPU.

BTW, only now read about Turin etc...was not paying attention to these future Epycs and whatnot before, last thing i knew was Zen4 Genoa 96C and Zen4C Bergamo 128C. So is Turin Zen5 = Genoa replacement and supposedly up to 192 cores, did i understand that right? And if yes, they are still keeping 8 cores per chiplet? Cause that would be 24 chiplets on one CPU...is it going to have 24 RAM channels? Or would 2 chiplets share single channel? Somehow i find this hard to believe to be truth...then again, what do i know.

 

[Joe NYC]

Meh. Give me 24 Zen5 cores not these hybrids. If i wanted that, i could get Intel CPU.

BTW, only now read about Turin etc...was not paying attention to these future Epycs and whatnot before, last thing i knew was Zen4 Genoa 96C and Zen4C Bergamo 128C. So is Turin Zen5 = Genoa replacement and supposedly up to 192 cores, did i understand that right? And if yes, they are still keeping 8 cores per chiplet? Cause that would be 24 chiplets on one CPU...is it going to have 24 RAM channels? Or would 2 chiplets share single channel? Somehow i find this hard to believe to be truth...then again, what do i know.

I think Turin will have 16 x 8 Core CCDs = 128
The Bergamo successor will likely have 12 x 16 core CCDs = 192

The socket stays the same, so the same 12 channels of memory.

 

[Timmah!]

I think Turin will have 16 x 8 Core CCDs = 128
The Bergamo successor will likely have 12 x 16 core CCDs = 192

The socket stays the same, so the same 12 channels of memory.

I see, that makes more sense. I read something about 192/256, so i thought 192 for Turin and 256 for Bergamo replacement.

 

[Geddagod]

Same thought, I only want the Zen 5 cores, not some puny outdated Zen 4c inebench cores.


I would guess one 8C Zen 5 Chiplet + 16C Zen 4 Chiplet for 24C.
So going by PHX2 topology, Zen 4c would have half the L3 per Core but for 16C in one CCX the amount of L3 would still be same? That could make L3 available to a core to the SW pretty much same even though L3 per core is half. But overall around same size as Zen 4 CCD but 2x cores, if going with PHX density of 25 MTr@178mm2
View attachment 78613

Zen 4C CCD is 50% bigger than the Zen 4 chiplet?
If Zen 5 uses that, I don't expect prices to be pretty for consumers
Especially since I think, at best, the Zen 5 chiplet is going to be the same size as the Zen 4 standard chiplet.
But also, wasn't there talk of Zen 4C using an even smaller cell type than Zen 4? Seeing how Zen 4 already uses HD libs, idk if AMD would go for even denser libs (or even if TSMC offers them as standard or if AMD would have to make their own libs), so idk how true this is.
Also, why use a Zen 4C chiplet anyway? Do we even know if Zen 4C will offer better MT perf/mm^2 or perf/cost than Zen 5? Over Zen 4 it's obvious, but for Zen 5 idk, since the node is similar so costs would be similar there, and Zen 5 should offer significant IPC gains... maybe if someone can do some rough napkin math for how much better Zen 5 would have to be to get better perf/watt over Zen4C...
If they do end up using Zen 4C though, hopefully we get a Zen 5 with V-cache and Zen 4C chiplet as an option...

 

[turtile]

Zen 5 is definitely not coming in 2023. AMD stated that it is a 2024 product. And since AMD refreshes their mobile line up every March, and has stated that the next mobile chip is based on Zen 5, we can conclude that Zen 5 should be out in March or April 2024.

The only thing I can see happening is AMD moving Zen 4 to a custom N4X for a November release. Zen 4 is smaller and the improved process can probably hit 6.2 Ghz. This extra clock speed would allow AMD to better compete with Intel in the mid to low-end retail space while Zen 5 takes the high end.

 

[DisEnchantment]

Amazing, were did you find that?
Slowly but surely all my wild speculations come true.
Next stop: IFoP successor 😎

From the manual for Phoenix and the generic architecture reference document.

The only thing I can see happening is AMD moving Zen 4 to a custom N4X for a November release. Zen 4 is smaller and the improved process can probably hit 6.2 Ghz. This extra clock speed would allow AMD to better compete with Intel in the mid to low-end retail space while Zen 5 takes the high end.

N4P should be enough for Zen 4 to reach 6.2 GHz. N4X is leaky and density is not great, from commercial perspective it will be not suitable for Zen 4.

Zen 4C CCD is 50% bigger than the Zen 4 chiplet?
If Zen 5 uses that, I don't expect prices to be pretty for consumers

PHX has a density of 140MTr/mm2. And the Zen 4 cores there hit 5.2 GHz. 16C Zen 4c with half L3 fabbed like this will be similar to a standard 8C Zen 4 chiplet in size, ~66 mm2.

Especially since I think, at best, the Zen 5 chiplet is going to be the same size as the Zen 4 standard chiplet.

That can't be the case.
From commercial perspective,
Zen 2 --> Zen 3 CCD got a 12% bump in area at a time when N7 is just in its 3rd year of production and AMD is unable to meet console SoCs demand and N7 supply is highly contested.
Zen 4 --> Zen 5 CCD should be >10% bump in area, when N5 family is in its 5th year of production (yes it is old) and additional AZ F21 fab online and in 2024 N5 wafer output will be close to double that of N7 in 2020 and other TSMC customers moving to N3 and AMD using N3, N5, N7 nodes at that point instead of just one.
From another perspective, Zen 4 CCD is very tiny, at 66.3 mm2 far below the optimal TSMC sweet spot of 80 mm2. Zen 3 is 80.7 mm2 partly because of this metric. N5 could actually improve this metric. It is 5 years old in 2024 after all.

Also, why use a Zen 4C chiplet anyway? Do we even know if Zen 4C will offer better MT perf/mm^2 or perf/cost than Zen 5? Over Zen 4 it's obvious, but for Zen 5 idk, since the node is similar so costs would be similar there, and Zen 5 should offer significant IPC gains... maybe if someone can do some rough napkin math for how much better Zen 5 would have to be to get better perf/watt over Zen4C...
If they do end up using Zen 4C though, hopefully we get a Zen 5 with V-cache and Zen 4C chiplet as an option...

Actually good point why Zen 4c, why not Zen 5c. But AMD differentiate efficiency cores from performance cores based on process efficiency mainly so they might not be ready to fab Zen 5c on N3E yet.

 

[BorisTheBlade82]

Zen4c will pretty surely have a better perf/mm2 than even Zen5. Almost doubling this KPI from OG Zen4 should sail them through this.

I would guess that the c-core line will be lagging the OG cores timeline-wise. For me it looks like this:

1. Zen4 in every segment
2. Zen4c in Bergamo and MI300
3. Zen4c in PHX2
4. Zen4c with Zen5 in the Desktop
5. Zen4c with Zen5 in mobile
6. Zen5c in DC
7. ...

 

[RTX2080]

Meh. Give me 24 Zen5 cores not these hybrids. If i wanted that, i could get Intel CPU.

BTW, only now read about Turin etc...was not paying attention to these future Epycs and whatnot before, last thing i knew was Zen4 Genoa 96C and Zen4C Bergamo 128C. So is Turin Zen5 = Genoa replacement and supposedly up to 192 cores, did i understand that right? And if yes, they are still keeping 8 cores per chiplet? Cause that would be 24 chiplets on one CPU...is it going to have 24 RAM channels? Or would 2 chiplets share single channel? Somehow i find this hard to believe to be truth...then again, what do i know.

It looks like AMD just only use Zen4C in its monolithic die like APUs which need to save place for mobile usage. When comes to dual CCD designs like Raphael it's close to impossible since the IO is limited to 8 core per CCD, and significantly contradict X3D dual CCDs design because AMD need high clock/X3D cores at the same time, while high density cores provide nothing.

My thought is AMD has many option but they don't need to, practically.

https://twitter.com/x/status/1636845531753316352

 

[naukkis]

When comes to dual CCD designs like Raphael it's close to impossible since the IO is limited to 8 core per CCD,

IO isn't limited to number of cores. Cores even don't access IO-die at all, they access L3 and when L3 misses it will make a request to IO-die. CCD of course can have more than one L3 like it was case with original Ryzen, there was two L3-domains that make request to integrated IO-part. If Zen4c-chiplet have shared L3 for all its cores it can be used with two chip configuration with Raphael IO too - other hand if it have two CCXs and two L3 clusters only one of such chip can be paired with Raphael IO-die which can only serve two L3-domains.

 

[BorisTheBlade82]

It is about certain that one Zen4c Bergamo CCD is housing 2 CCX with 8 cores each. These two CCX do not share their L3, just as was the case with Zen2.

Regarding the interconnect to the IOD the picture is much more blurry:

• One option is: The current EPYC IOD has 12 IFoP links. So the up to 8 Bergamo CCD could get connected via one link each. That could get them bandwidth starved as this is only half the bandwidth per core as Desktop and Server Zen4 have.
• Another option: From my measurements it seems, that MI300 employs 24 Zen4c cores, not Zen4. One die could be the plain old 16c die for Bergamo. And that might be connected via Infinity Fan-out Links AKA InFO-RDL. This would bring them plenty bandwidth even with a rather narrow link. So for me it is quite possible, that starting with Bergamo and later on with Desktop Zen5 AMD will get rid of the current IFoP.

 

[Kepler_L2]

It is about certain that one Zen4c Bergamo CCD is housing 2 CCX with 8 cores each. These two CCX do not share their L3, just as was the case with Zen2.

Regarding the interconnect to the IOD the picture is much more blurry:

• One option is: The current EPYC IOD has 12 IFoP links. So the up to 8 Bergamo CCD could get connected via one link each. That could get them bandwidth starved as this is only half the bandwidth per core as Desktop and Server Zen4 have.
• Another option: From my measurements it seems, that MI300 employs 24 Zen4c cores, not Zen4. One die could be the plain old 16c die for Bergamo. And that might be connected via Infinity Fan-out Links AKA InFO-RDL. This would bring them plenty bandwidth even with a rather narrow link. So for me it is quite possible, that starting with Bergamo and later on with Desktop Zen5 AMD will get rid of the current IFoP.

The CCDs in MI300 are on top of the base I/O die and connected via SoIC.

 

[BorisTheBlade82]

The CCDs in MI300 are on top of the base I/O die and connected via SoIC.

Right, they are connected to a base die via SoIC. But how are the four base dies connected to HBM and quite likely each other as well? The most likely answer after having a look at N31 is InFO-RDL. Another option would be EFB for connecting the base dies to each other.

 

[DisEnchantment]

But how are the four base dies connected to HBM and quite likely each other as well?

Not InFO-RDL very likely , the BW requirements for the GPU SEs is massive. LSI and/or EFB.

 

[BorisTheBlade82]

Not InFO-RDL very likely , the BW requirements for the GPU SEs is massive. LSI and/or EFB.

Well, for the MCD in N31 we already have 1 TByte/s. This can be increased quite easily by using more beachfront.
Apple's M1 Ultra, which employs InFO-LSI or EFB also has only 2.5 Tbyte/s.

/edit:
But in general I agree with you. A silicon bridge is more likely.

Apple Announces M1 Ultra: Combining Two M1 Maxes For Workstation Performance

www.anandtech.com

 

[RTX2080]

Gigabyte's spokesman must be trembling right now.

Gigabyte claims AMD will release “next generation” Ryzen desktop CPUs this year - KitGuru

While AMD hasn't announced its next-gen Ryzen processor plans yet, it would seem that partners are a

www.kitguru.net

Rumor: Next-Generation AMD Ryzen 8000 Desktop Processors: A Surprise Release in 2023?

In an unexpected turn of events, hardware manufacturer Gigabyte inadvertently revealed that the next generation of AMD Ryzen desktop processors may be released later this year....

www.guru3d.com

Gigabyte reportedly confirms AMD Ryzen 8000 release window: Performance improvements, availability, and more

Gigabyte has reportedly confirmed that the next-gen AMD Ryzen 8000 CPUs will arrive later this year.

www.sportskeeda.com

GIGABYTE Stirs Zen 5 and Ryzen 8000 Series Speculation: "Next Generation of AMD Ryzen Desktop CPUs Will Come Out Later This Year"

AMD is planning to release a new generation of Ryzen desktop processors, and while that isn't much of a surprise in itself, it could be happening sooner than many might have expected. Giga Computing, a subsidiary of GIGABYTE and an industry leader in high-performance servers, server...

www.thefpsreview.com

AMD Ryzen 8000: Neue Generation soll laut Gigabyte noch dieses Jahr erscheinen

Laut Gigabyte will AMD noch dieses Jahr eine neue Generation Ryzen-CPUs auf den Markt bringen.

www.pcgameshardware.de

 

[Doug S]

Well, for the MCD in N31 we already have 1 TByte/s. This can be increased quite easily by using more beachfront.
Apple's M1 Ultra, which employs InFO-LSI or EFB also has only 2.5 Tbyte/s.

Apple is using a massive number of I/Os clocked fairly slowly to minimize power consumption. It is the result of a different set of goals.

 

[BorisTheBlade82]

Apple is using a massive number of I/Os clocked fairly slowly to minimize power consumption. It is the result of a different set of goals.

So you are saying, they may have decreased TSMC's pj/bit figures by lowering the GByte/mm as well through lower clocks?

 

[Doug S]

So you are saying, they may have decreased TSMC's pj/bit figures by lowering the GByte/mm as well through lower clocks?

Yes, similar to how you can reduce watts per unit of performance by undervolting & underclocking a CPU. We know Apple is using on the order of 10,000 I/Os and the data rate is 2.5 TB/sec i.e. ~2 Gbps per pin. So it should be easy to calculate the specifics and compare if TSMC has released the required detail.

 

[A///]

Zen 5 is definitely not coming in 2023. AMD stated that it is a 2024 product. And since AMD refreshes their mobile line up every March, and has stated that the next mobile chip is based on Zen 5, we can conclude that Zen 5 should be out in March or April 2024.

The only thing I can see happening is AMD moving Zen 4 to a custom N4X for a November release. Zen 4 is smaller and the improved process can probably hit 6.2 Ghz. This extra clock speed would allow AMD to better compete with Intel in the mid to low-end retail space while Zen 5 takes the high end.

as i suspected it's another xt. sit back and wait for @DrMrLordX to complain about the xts again.

 

[DrMrLordX]

as i suspected it's another xt. sit back and wait for @DrMrLordX to complain about the xts again.

Nah I'll let @Kocicak do it this time.

 

[Kocicak]

I dont understand what should I complain about...

Seeing the post by Turtile

This extra clock speed would allow AMD to better compete with Intel in the mid to low-end retail space while Zen 5 takes the high end.

I can just remark that thanks to Intel flooding the CPUs with E cores it is very hard to compete with them in the low/mid end market segment.

 

[A///]

I dont understand what should I complain about...

Nothing, he doesn't want to admit he's fussy about AMD's release schedule. the pandemic messed up their scheduling. IIRC Zen 4 was later than expected at 24 months. Zen 5 returns to the 14-18 month schedule they had prior.

 

[Tigerick]

There is new leaks from RGT about Strix Point, AMD seems like going to split the APU design into 3 die, below is comparison table:

Name	Model	Launch Date	Node	CPU cores	L3 Cache	Memory LPDDR5x	Memory BW	GPU	ALU	IC	TDP
PHX2	R5 7040	Q3 2023	N4 115mm2	2xZen4 + 4xZen4c	4+4=8MB	64-bit 7500	60 GB/s	RDNA3 4CU 256SP	256	NA	15W
STX3	R5 8050	Q3 2024	N4P	2xZen5 + 4xZen5c	12 MB	64-bit 8533	68 GB/s	RDNA3+ 4CU 256SP	512	NA	15W
PHX+	R7&9 8040	Q1 2024	N4P	8xZen4	16 MB	128-bit 8533	136 GB/s	RDNA3+ 12CU 768SP	1536	NA	15-45W
STX2	R7 8050	Q3 2024	N4P	4xZen5 + 8xZen5c	24 MB	128-bit 8533	136 GB/s	RDNA3+ 8CU 512SP	1024	NA	15-45W
STX1 Halo	R9 8050	Q3 2024	N4P + N4P	8xZen5 + 8xZen5c	32 MB	256-bit 8533	272 GB/s	RDNA3+ 20CU 1280SP	2560	32MB	25-120W
				6xZen5 + 8xZen5c	28 MB	128-bit 8533	136 GB/s	RDNA3+ 10CU 640SP	1280	32MB
Fire Range	R7&9 8055	Q3 2024	N4P x 3	16xZen5	64 MB	128-bit		RDNA3+ 2CU 128SP	256	NA	55W+

Guess AMD still going monolithic with Strix Point with 3 different dies

 

[Joe NYC]

There is new leaks from RGT about Strix Point, AMD seems like going to split the APU design into 3 die, below is comparison table:

PHX2	N4 115mm2	2xZen4 + 4xZen4c	8MB	RDNA3 4CU 256SP	15W
STX3	N4P	2xZen5 + 4xZen4c	8MB	RDNA3+ 4CU 256SP	15W
STX2	N4P	4xZen5 + 8xZen4c	16MB	RDNA3+ 8CU 512SP	~30W?
STX1	N3E	8xZen5 + 8xZen4c	32MB	RDNA3+ 16CU 1024SP	~35W?

Guess AMD still going monolithic with Strix Point with 3 different dies

Kind of boring, I have to say.

The STX1 would seem quite big to be a monolithic die.

It would be surprising if AMD ends up with ~250+ mm2 monolithic die to go against chiplet based Meteor Lake and Arrow Lake. I don't see this as likely...