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

[DisEnchantment]

Speculate at will

 

[yuri69]

Zen 5’s Leaked Slides

A YouTuber called Moore’s Law is Dead recently leaked a couple AMD slides about Zen 5.

chipsandcheese.com

Some real good stuff here, including info from Zen4 on current "bounds" of improvement.
While i personally think diminishing returns will apply strongly on Z5, i think it should be good for 15+% IPC AND open up further optimization of 10+% IPC on Z6.

Intel's been throwing massive resources at Sunny-Golden Coves and managed ~19% IPC gains in both cases. Yet, AMD is hitting diminishing returns. OK

 

[JoeRambo]

Intel's been throwing massive resources at Sunny-Golden Coves and managed ~19% IPC gains in both cases.

19% is quite nice actually, Apple has spent considerable resources and has not improved IPC much with latest gen or two?

Yet, AMD is hitting diminishing returns. OK

Only a true ADF member can offer example supporting case for diminishing returns ( even if from different vendor ) and proceed to claim it won't impact AMD. All in same line of text.

 

[Geddagod]

Intel's been throwing massive resources at Sunny-Golden Coves and managed ~19% IPC gains in both cases. Yet, AMD is hitting diminishing returns. OK

GLC was a 15% IPC uplift in server vs SNC. Plus GLC was esentially just a larger SNC, if Zen 5 goes massively wider, this is a pretty significant change from AMD's previous strategy of going smaller structure sizes with low latency vs large structures with higher latency. Even Zen 3 doesn't look to be as big of a change of structure size vs Zen 2, compared to what we might be getting with Zen 5 vs Zen 4.

 

[eek2121]

Intel's been throwing massive resources at Sunny-Golden Coves and managed ~19% IPC gains in both cases. Yet, AMD is hitting diminishing returns. OK

While I am not disagreeing with the article or any of the opinions/discussions here, you are comparing 2 completely different processors from two very different companies.

Zen 4, despite being a much smaller core, has no issue keeping up with Golden Cove. AMD has taken the Zen architecture and refined/optimized it over several generations. This alone leads me to believe the gains will be smaller for Zen 5. I do hope I am wrong, of course. An overall 25-35% uplift at same power would be an instabuy for me. I suspect they won’t come close, however. Likely, total performance uplift will range from 3%-20% depending on workload.

Remember client Zen 5 is going to be on N4P. The differences between N4P and the custom N5 spin that AMD uses aren’t significant. AMD will not be looking to increase die area (or they would be sacrificing margins). That leaves little room for improvement.

 

[igor_kavinski]

That leaves little room for improvement.

AMD must have designed Zen5 to target ST domination of Arrow Lake for winning over the PC gaming crowd (both Intel and AMD get wind of future CPU performance projections from moles/insiders in Taiwanese companies. That's the only way the CPU race ends up being so close). So considering that bit of speculation on my part, Zen 5 must have enough architectural enhancements to beat 2024 Arrow Lake by minimum 10% in ST. Can't say anything about MT coz much clouded the future is. Intel may counter Zen 5 with their 40-core Beast Lake in H1 2025. The reason I think Arrow Lake may end up staring at the rear end of Zen 5 is Intel's team of monkey engineers being hamstrung by the huge size of their P-core and process related woes such as not having enough time to refine their nodes.

I don't think most enthusiasts will have the patience to wait several months more for Arrow Lake's launch when they can get a really performant Zen CPU in H1 2024.

 

[Thunder 57]

AMD must have designed Zen5 to target ST domination of Arrow Lake for winning over the PC gaming crowd (both Intel and AMD get wind of future CPU performance projections from moles/insiders in Taiwanese companies. That's the only way the CPU race ends up being so close). So considering that bit of speculation on my part, Zen 5 must have enough architectural enhancements to beat 2024 Arrow Lake by minimum 10% in ST. Can't say anything about MT coz much clouded the future is. Intel may counter Zen 5 with their 40-core Beast Lake in H1 2025. The reason I think Arrow Lake may end up staring at the rear end of Zen 5 is Intel's team of monkey engineers being hamstrung by the huge size of their P-core and process related woes such as not having enough time to refine their nodes.

I don't think most enthusiasts will have the patience to wait several months more for Arrow Lake's launch when they can get a really performant Zen CPU in H1 2024.

Surely you know the PC gaming crowd is an afterthought at best, right? They will do what they think is best for servers. If that happens to help gamers, so be it.

 

[Saylick]

While I am not disagreeing with the article or any of the opinions/discussions here, you are comparing 2 completely different processors from two very different companies.

Zen 4, despite being a much smaller core, has no issue keeping up with Golden Cove. AMD has taken the Zen architecture and refined/optimized it over several generations. This alone leads me to believe the gains will be smaller for Zen 5. I do hope I am wrong, of course. An overall 25-35% uplift at same power would be an instabuy for me. I suspect they won’t come close, however. Likely, total performance uplift will range from 3%-20% depending on workload.

If Zen 5 is an iteration of Zen 4 then I would agree, but so far it seems that Zen 5 is supposed to be a ground-up rewrite of the core, a la Zen 3, so I don't see why Zen 5 would have smaller gains than Zen 4. There's too many changes under the hood for Zen 5 to basically be an enhanced Zen 4 imo. With a much wider front end, one must rebalance the entire core to get the most out of it. At that point, you've essentially done a rebuild.

Remember client Zen 5 is going to be on N4P. The differences between N4P and the custom N5 spin that AMD uses aren’t significant. AMD will not be looking to increase die area (or they would be sacrificing margins). That leaves little room for improvement.

If the performance improvement outweighs the increase in die area, then it's definitely worth it imo. Also, it is worth pointing out that on a Zen 4 CCD roughly half the die is L3 cache, so even with a 50% larger core, it only amounts to a 25% larger CCD (assuming L3 cache size doesn't change). If that 50% larger core gives you >25% more performance then it's worth it.

 

[igor_kavinski]

They will do what they think is best for servers.

They don't really have to try that hard.

Intel *cough* Xeons *cough* have been stellar so far.

/s

 

[adroc_thurston]

AMD will not be looking to increase die area (or they would be sacrificing margins)

Yes they are lol.
The core targeted N3 before it exploded.
Throwing xtors at the problem was always the idea.

Apple has spent considerable resources

que?
It's a skeleton crew.

 

[AMDK11]

Plus GLC was esentially just a larger SNC,

I cannot agree with this. The core of GoldenCove is not only a slightly larger SunnyCove. GoldenCove is a significantly expanded and rebuilt new x86 core, and the number of transistors in the core logic has increased similarly to SunnyCove (+38%) compared to Skylake.

 

[Geddagod]

I cannot agree with this. The core of GoldenCove is not only a slightly larger SunnyCove. GoldenCove is a significantly expanded and rebuilt new x86 core, and the number of transistors in the core logic has increased similarly to SunnyCove (+38%) compared to Skylake.

I didn't even say "slightly larger", and yes, GLC is esentially a larger SNC. It's not even close to a clean sheet design like Zen was, or netburst, or nehalem iirc. Oh sure, Intel is gonna call GLC the "architecture for the next decade of compute", but it's not anything so special lol. The only way it's going to be the next architecture of the decade is if ARL gets delayed any further lmao.
Besides, how much more semantical can this get? That was obviously not even my point... Zen 5 vs Zen 4 appears to be a change not just in "haha core goes wide", but literally AMD's entire design philosophy- away from narrow and fast to wider designs. Intel on the other hand has been doing "haha core goes wide" since SNC.
But sure, if it's really that deep, Golden Cove is a new core lol.

The core targeted N3 before it exploded.

Common TSMC and AMD L

They don't really have to try that hard.

Intel *cough* Xeons *cough* have been stellar so far.

/s

Queue Nvidia claiming they chose SPR for the stronger per-core performance
Honestly stronger 1T perf is prob very useful for servers as well, but lets not pretend AMD cares about gamers lol

Remember client Zen 5 is going to be on N4P. The differences between N4P and the custom N5 spin that AMD uses aren’t significant. AMD will not be looking to increase die area (or they would be sacrificing margins). That leaves little room for improvement.

Eh I disagree, AMD increased core area with Zen 3, so it's not like they haven't increased core area before.

 

[Timmah!]

If Zen5 on AM5 will be limited to 16C with ~15% perf increase while bumping price it'll be a dud.

Unless we're getting a Zen5 core count increase to 24/32C (whether through 24/32C or 16P+8/16E), I think buying Zen4 at Black Friday at a discount will be a much better option. Intel 14xxx series will also have been released by that time, putting even further price pressure on Zen4. Currently 7950X3D is around $650, so are we expecting roughly $499 at BF?

If i were to buy new machine, i guess i would rather wait for Zen5, unless i was budget constrained and would be forced to take cheaper option in form of discounted Zen4.
As an owner of Zen4 however, 16C with 15 percent more performance for higher price? I am not sure thats worth for the same price as before.

 

[igor_kavinski]

Honestly stronger 1T perf is prob very useful for servers as well, but lets not pretend AMD cares about gamers lol

They are the only ones giving gamers a huge cache with the 7800X3D. Gamers are an important demographic for them. If it seems they could do more, yes but they have much more limited resources than Intel and they have done way more than Intel with what they got.

 

[adroc_thurston]

They are the only ones giving gamers a huge cache with the 7800X3D

It's a direct Genoa-X offshoot.

 

[H433x0n]

They are the only ones giving gamers a huge cache with the 7800X3D. Gamers are an important demographic for them. If it seems they could do more, yes but they have much more limited resources than Intel and they have done way more than Intel with what they got.

AMD treats gamers like plebs. They literally get table scraps in a suboptimal packaging solution. Edit - I mean sub optimal for client desktop.. not necessarily server.

Luckily this appears to be changing with Zen 6 and client will get a proper dedicated design.

 

[AAbattery]

SMH

You are just trying to justify a purchase coz you can't wait.

Remember, this quote is from a CPU architect at the top of his game: https://www.anandtech.com/show/17031/anandtech-interviews-mike-clark-amds-chief-architect-of-zen


He did not say anything remotely close to that for Zen 4.

Maybe Mike was excited more by the interesting changes in architecture than purely by the performance improvement versus the previous generation.

Zen 4 got a lot of its performance increase from clock speed, which might not seem as exciting from the technical perspective of a CPU architect. Meanwhile Zen 5 is rumored to get all its performance increases from architecture versus clock speed.

Luckily this appears to be changing with Zen 6 and client will get a proper dedicated design.

Except it will be a compromise between laptop and desktop, favoring the needs of the laptop segment.

 

[adroc_thurston]

Luckily this appears to be changing with Zen 6 and client will get a proper dedicated design.

Well ughhh.
Kinda?

 

[Thunder 57]

AMD treats gamers like plebs. They literally get table scraps in a suboptimal packaging solution. Edit - I mean sub optimal for client desktop.. not necessarily server.

Luckily this appears to be changing with Zen 6 and client will get a proper dedicated design.

We still don't know everything about Zen 5, and Zen 6 info is already out? EPYC and Ryzen are diverging? Interesting choice.

 

[H433x0n]

Well ughhh.
Kinda?

Wat.

I thought you said it’d be utilizing the latest CoWoS packaging?

 

[adroc_thurston]

I thought you said it’d be utilizing the latest CoWoS packaging?

InFO but that doesn't make it a dedicated desktop chip.
Or that it would make it in any way, shape or form better for desktop.
It's just higher bandwidth links at lower power; i.e. made for mobile.

 

[itsmydamnation]

InFO but that doesn't make it a dedicated desktop chip.
Or that it would make it in any way, shape or form better for desktop.
It's just higher bandwidth links at lower power; i.e. made for mobile.

Can you give us anymore detail how you got to 32% IPC. is it extrapolated or a direct result ?

i think the people who are going 15% IPC is good are on full on hopium ( just look at their history).
If Zen5 is 15% IPC increase then its a failure of a core given the resource increase and that's my reverse hopium .

 

[adroc_thurston]

Can you give us anymore detail how you got to 32% IPC. is it extrapolated or a direct result ?

AMD gave projected SIR2017 scores for Turin/Turin-D when they did the sampling thingy.
Those turned out to be a wee bit higher than I expected.

 

[TESKATLIPOKA]

Considering AMD didn't choose the option to increase the core count above 16 cores, I think the performance per core should see a good enough increase.

It's true that Strix point did see a 50% increase in core count, but 2/3 are dense ones, so the area should be very close between 8*Zen5 vs 4*Zen5+8*Zen5c and Strix is aimed for lower TDP, so this combination should provide better perf/W.

 

[AMDK11]

I didn't even say "slightly larger", and yes, GLC is esentially a larger SNC. It's not even close to a clean sheet design like Zen was, or netburst, or nehalem iirc. Oh sure, Intel is gonna call GLC the "architecture for the next decade of compute", but it's not anything so special lol. The only way it's going to be the next architecture of the decade is if ARL gets delayed any further lmao.
Besides, how much more semantical can this get? That was obviously not even my point... Zen 5 vs Zen 4 appears to be a change not just in "haha core goes wide", but literally AMD's entire design philosophy- away from narrow and fast to wider designs. Intel on the other hand has been doing "haha core goes wide" since SNC.
But sure, if it's really that deep, Golden Cove is a new core

.
I think that when it comes to GoldenCove being the basis of the generation for the next decade, it is about 6-way x86 decoding.In my opinion, LionCove, although it has also been greatly expanded, is also based on a 6-way x86 decoder.I am very curious how much it has been expanded and whether they will be able to maintain the average IPC increase at the level of 18-19%.

 

[DisEnchantment]

If the performance improvement outweighs the increase in die area, then it's definitely worth it imo. Also, it is worth pointing out that on a Zen 4 CCD roughly half the die is L3 cache, so even with a 50% larger core, it only amounts to a 25% larger CCD (assuming L3 cache size doesn't change). If that 50% larger core gives you >25% more performance then it's worth it.

In a Zen 4 CCD of 66mm2 (w/o scribe lines),

• 17% (11.22mm2) is GMI+DBG+DFT+SMU+TSV out of 66mm2
• 58% (32mm2) is L2+L3 out of 55mm2
• 42% (23mm2) is Core out of 55mm2

Zen 2@72mm2 --> Zen 3@80.7mm2 --> Zen 4@66mm2

Let's say a minor bump of 12% CCD size, Zen 4@66mm2 --> Zen 5@74mm2
That is a 35% increase in core area (with L2/L1 remaining largely same) 600+ MTr per core

A 20% bump in CCD size, i.e. 80mm2 Zen 5 CCD would lead to 60% bigger core.