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

[DisEnchantment]

Speculate at will

 

[CakeMonster]

If Zen5 X3D will be released already in September, are any of you intending to buy the vanilla variant in July anyway and if so why?

I want the 9950X, and prefer to avoid the thread prioritization mess of the 2 CCD X3D models. It's also available earlier, and is cheaper, and more likely to work on W10.

And while I guess it could happen, I would not count on the X3D lineup being available in september.

 

[DisEnchantment]

Yep. Thing is, parameterization of stuff like that often costs more in dev/validation than you save in an incrementally smaller core variant. AMD has been wise to mostly avoid it thus far.

I think anyone believing that Strix Zen 5 will somehow be a materially worse microarchitecture than desktop Zen 5 is going to be disappointed.

Leaked slides already showed different FP width variants, why not?

 

[SarahKerrigan]

Leaked slides already showed different FP width variants, why not?
View attachment 101542

Different animal. Means absolutely nothing for PRF. Do you have any evidence whatsoever that Strix core structure sizes are smaller than desktop Zen5?

Looking forward to "sure, Strix ST isn't a huge boost, but my +32% is totally coming on desktop!"

 

[DisEnchantment]

Looking forward to "sure, Strix ST isn't a huge boost, but my +32% is totally coming on desktop!"

Take the 32% with someone else not me, not interested in any debate.

Different animal. Means absolutely nothing for PRF. Do you have any evidence whatsoever that Strix core structure sizes are smaller than desktop Zen5?

In case you did not notice I said possibly, because I don't state anything as fact without manuals or patches

 

[fastandfurious6]

why cant they make a top-end CPU that is pure CPU only? i.e.

- no NPU
- no GPU (not even 2CU)
- no AVX-512 (variants maybe?)
- etc

all the die area for raw compute power

why not?

 

[CouncilorIrissa]

why cant they make a top-end CPU that is pure CPU only? i.e.

- no NPU
- no GPU (not even 2CU)
- no AVX-512 (variants maybe?)
- etc

all the die area for raw compute power

why not?

Because it would be a desktop-only solution, i.e. DOA.

• You need an NPU to get the Copilot+ sticker;
• You need the iGPU for thin&light laptops to drive down the costs and power;
• You need AVX-512 for server at the very least.

Desktop is a niche market.

edit: not even a desktop-only solution actually, a Gamer Desktop™. You can get away with an iGPU on desktop if you're using the PC as a workstation.

 

[SarahKerrigan]

• You need AVX-512 for server at the very least.

For supercompute stuff, sure, but that's a tiny percentage of the market.

I've not see a lot of indication that the meat of the server market cares much about AVX (or even much about floating-point.) Maybe AI is changing that, but if so, it's slow going.

 

[CouncilorIrissa]

For supercompute stuff, sure, but that's a tiny percentage of the market.

I've not see a lot of indication that the meat of the server market cares much about AVX (or even much about floating-point.) Maybe AI is changing that, but if so, it's slow going.

Thanks for the correction, I'm not exactly an expert in the server market. I derived my conclusion mostly from Intel debuting the AVX-512 in SKL-X. Any clue as to what market they intended to capture?

 

[SarahKerrigan]

Thanks for the correction, I'm not exactly an expert in the server market. I derived my conclusion mostly from Intel debuting the AVX-512 in SKL-X. Any clue as to what market they intended to capture?

HPC and technical computing stuff. It exists, it's not something worth disregarding, but it's not like a huge percentage of deployed cores are just running vector operations day-in and day-out. A lot of server stuff just looks like the same sort of integer spaghetti that's common on desktops.

AMD, I think, has generally understood this well. They have not been as aggressive as Intel on vector resources, and it's worked out fine for them, because the users that really deeply care are a niche.

 

[jpiniero]

AMD, I think, has generally understood this well. They have not been as aggressive as Intel on vector resources

The obsession with AI might be changing that.

 

[SarahKerrigan]

The obsession with AI might be changing that.

Of course. I'm not qualified to examine that in any significant depth because I don't have much experience with, or interest in, AI.

 

[gdansk]

The obsession with AI might be changing that.

If that's why it is a true misapplication. VNNI will be completely outclassed by any proper matrix extension like AMX.

FP512 seems more fit for supercomputer contracts. But for whatever reason all the CCD based products get it.

 

[Fjodor2001]

I want the 9950X, and prefer to avoid the thread prioritization mess of the 2 CCD X3D models. It's also available earlier, and is cheaper, and more likely to work on W10.

And while I guess it could happen, I would not count on the X3D lineup being available in september.

Well, what would be the practical problem of 2 CCD X3D on Zen5? Are you expecting it to perform worse than the corresponding vanilla variant?

And what problems are you expecting with Zen5 X3D on Win11/12?

 

[inquiss]

If that's why it is a true misapplication. VNNI will be completely outclassed by any proper matrix extension like AMX.

FP512 seems more fit for supercomputer contracts. But for whatever reason all the CCD based products get it.

Since we have the next gen zen 6 mobile and desktop having the same CCD, rather than desktop and server up until now, I think we could have a situation where zen 6 desktop chips don't have AVX 512 as that server only. This could be the only desktop gen that wide.

 

[Jayzen]

I've been a pessimist, but I'm surprised that Zen 5 actually has 5% worse IPC than Zen 4 in many places according to that credible leak. Pair that with the expected clock regression due to a wider core and...yeah. This is shaping up to be just as bad as Bulldozer, complete with being consistently outperformed by last generation's product.

 

[Markfw]

For supercompute stuff, sure, but that's a tiny percentage of the market.

I've not see a lot of indication that the meat of the server market cares much about AVX (or even much about floating-point.) Maybe AI is changing that, but if so, it's slow going.

OK, first, as said, AI may be changing that, since I have read that avx-512 is quite useful there. Next, it may be niche, but in the DC community, just in ONE projects of hundreds, There are almost 1000 users that participated in the last event alone, and many of those have 10-20 PCs , and avx-512 is very useful for this project. Even if all 1000 had 20 PCs, all Zen 4 based, I realize that is only 20,000 desktops, but if you multiply that by the other projects, this alone may be 100,000 PC in the market for it, worldwide. I think thats a little more than a tiny percentage. That not not included workstations that may benefit from it. My numbers may not be exactly accurate, but again, its a little more than a tiny percentage.

 

[SarahKerrigan]

I've been a pessimist, but I'm surprised that Zen 5 actually has 5% worse IPC than Zen 4 in many places according to that credible leak. Pair that with the expected clock regression due to a wider core and...yeah. This is shaping up to be just as bad as Bulldozer, complete with being consistently outperformed by last generation's product.

This is silly. Any major microarchitectural change has outlier code streams that don't like it. Zen 5 is not shaping up to be any kind of a disaster.

I'd remind you that Bulldozer was lagging Intel competition by 20%+ for ST int.

 

[Doug S]

This is silly. Any major microarchitectural change has outlier code streams that don't like it. Zen 5 is not shaping up to be any kind of a disaster.

I'd remind you that Bulldozer was lagging Intel competition by 20%+ for ST int.

Zen 5 is going to destroy Intel and Apple, it will be a 40%+ IPC gain across the board!

Zen 5 is a disaster, it is as bad as Bulldozer!

How quickly the wind changes direction around here

 

[CouncilorIrissa]

This is an obvious bait.

How does one derive “Z5 is 5% slower than Z4 iso clock on average” from “Zen 5 iso clock performance ranges from -5% to +24% compared to Zen 4” otherwise.

 

[Markfw]

OK, first, as said, AI may be changing that, since I have read that avx-512 is quite useful there. Next, it may be niche, but in the DC community, just in ONE projects of hundreds, There are almost 1000 users that participated in the last event alone, and many of those have 10-20 PCs , and avx-512 is very useful for this project. Even if all 1000 had 20 PCs, all Zen 4 based, I realize that is only 20,000 desktops, but if you multiply that by the other projects, this alone may be 100,000 PC in the market for it, worldwide. I think thats a little more than a tiny percentage. That not not included workstations that may benefit from it. My numbers may not be exactly accurate, but again, its a little more than a tiny percentage.

To add. For this one project, here are the total participating users and computers:

User Count

355 157

Host Count

859 486

so, over 800k computers for this one project that can seriously use avx-512.
And the number of projects that can use avx-512 is up to 100. 10 of which I am sure of.

So a good chance that 10 million PC's for this area.

Anybody want to estimate how many server chips could use it ?

 

[eek2121]

why cant they make a top-end CPU that is pure CPU only? i.e.

- no NPU
- no GPU (not even 2CU)
- no AVX-512 (variants maybe?)
- etc

all the die area for raw compute power

why not?

A GPU is useful for a few reasons. It was a welcome addition to Zen 4.

You also do not appear to understand how the functionality listed above helps regular users.

The onboard GPU can be used by Windows to accelerate things like virus scanning. A hybrid mode can also be utilized to save power, that is, Windows can use the APU until a demanding application is run.

Some ways the NPU can be used (not limited to this list):

1. Removing or blurring the background of a photo, or even entire objects.
2. Enhanced local/offline language translation, including OCR support.
3. For gamers, vendor independent upscaling, enhanced game AI, and much more.
4. Enhanced text to speech, speech to text (dictation)
5. Anything that benefits from machine learning, like auto photo/file grouping.
6. Enhanced audio playback, even if/when the source is poor quality.

AVX-512 likewise accelerates certain workloads. It hasn’t seen more adoption because it usually requires developers do some manual optimization, but when it is used, the uplift is big.

NOTE: Desktop Zen 4 will NOT have an NPU.

EDIT: As node shrinks slow down/hit theoretical limits, specialized instruction sets like AMX, AVX-512, etc. will be needed to continue to boost performance. This includes yet uninvented or unreleased instruction sets. I also speculate we may see a return to CMT for this reason.

 

[fastandfurious6]

Because it would be a desktop-only solution, i.e. DOA.

• You need an NPU to get the Copilot+ sticker;
• You need the iGPU for thin&light laptops to drive down the costs and power;
• You need AVX-512 for server at the very least.

Desktop is a niche market.

edit: not even a desktop-only solution actually, a Gamer Desktop™. You can get away with an iGPU on desktop if you're using the PC as a workstation.

no, I'm 100% sure there's a major market for lightweight laptop + raw compute power

i.e. yes the main model can be npu+gpu etc for the reasons you said

BUT

why don't they just make a variant that is raw compute power for laptops? and the whole die optimized for maximum CPU power
it will champion every single chart (except ST I guess but close enough)

i.e.

strix halo will have a whopping 40CUs + 16 full cores + NPU etc

theoretically without CUs or NPU or AVX (unnecessary on laptops!), it could have 32 cores no?

 

[CouncilorIrissa]

What you’re describing is basically Dragon Range.

Look up how many designs it has (spoiler: it’s like one laptop).

So yeah, it’s DOA.

 

[fastandfurious6]

but no it still baffles me

halo has 16full cores + 40cus (!!!!!!!!) + npu

zen 5 fire range will have just 16 cores, basic 2cu no npu

what

 

[static shock]

The desktop 8C/16T version looks attractive. 65W for better than todays AMD/Intel solutions is pretty nice. AMD crammed things on Zen4 to dig double digits performance uplift without turn Z5 a power hog. But 3nm is where should be Nirvana todays. No mention that AMD could wide it even more to dig scary fast performance uplifts(to make zen6 a little fatty but they can).

EDIT: let for we get AVX512 for some specific loads(ps3 emulation) that us power user want/need!

Let zen6 to be wider too AMD, Intel is throuwing caution to the wind with ARL. AMD can do it on 3N/2N-GAA without power consumption caution.

EDIT2: Intel LNC is 9-wide issue and it is a beast(by a good way)!