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

[DisEnchantment]

Speculate at will

 

[Timmah!]

No, much the same way -halo mobile isn't b.L.
You can pay for real cores.

I am not fan of b.L., so glad to hear that.
Then again not a fan of mobile chip replacing desktop. If i would be looking to get mobile level of performance, then i would be buying mobile device, not desktop.

 

[adroc_thurston]

i would be looking to get mobile level of performance, then i would be buying mobile device, not desktop.

Is Strix-Halo not enough CPU power?

 

[gdansk]

Is Strix-Halo not enough CPU power?

No one knows! (Except you)

 

[adroc_thurston]

No one knows! (Except you)

Well it's 16c Zen5.

 

[gdansk]

Well it's 16c Zen5.

Yes but enough is relative & for the non-psychics its performance relative to Intel competition remains to be seen.

 

[JustViewing]

AutoCAD, 3Dsmax, stuff like that.

Assuming both will cost around 999$, if you are a user of those applications and given the choice between 16 core Z5 vs 32 core Z5, which one would you choose? Would a 30% increase in single thread performance more important than near double performance for multi threaded applications?

 

[JustViewing]

Also, AMD stated only 16 core for Zen 5 already I believe so that is moot.

Did AMD mentioned that? Can you link to it? Even if they did, AMD is open to suggestions. Their suggestion box is open to all employees, if there is enough interest they may go beyond 16 cores. Lisa mentioned that she reads tech forums, if she sees enough enthusiasm for 16+ cores, she has the power to implement it .

 

[Det0x]

Hey Lisa, i dont want two CCD's with 16 Zen5c cores each on a mainstream desktop part
Give me dual CCD with 8 Zen5 for highest possible ST performance together with a doubling of the GMI links so each CCD can use higher than ~70GB/s max

Zen4 dual CCD memory performance maxed out:

Zen4 single CCD memory performance maxed out:


*edit*
More readable single CCD numbers:

 

[A///]

@Joe NYC will hurt you, detox.

 

[adroc_thurston]

relative to Intel competition remains to be seen

Intel isn't competitive, ARL-S is a 5% ST bump.
Better luck next time!

together with a doubling of the GMI links

Just wait for Zen6 where GMI goes poof or buy an STX-Halo laptop where it also goes poof.

 

[Gideon]

Hey Lisa, i dont want two CCD's with 16 Zen5c cores each on a mainstream desktop part

Give me dual CCD with 8 Zen5 for highest possible ST performance together with a doubling of the GMI links so each CCD can use higher than ~70GB/s max

Yeah, while i know this is unlikely to happen, I would also like to see a "universal halo-product" Zen 5 version:

Ryzen 9 8990X3D that has the following config:

• 1st CCD - 8c Zen 5 CCD with 3D Cache
• 2nd CCD - 16c Zen 5c

Imo this would be the CPU best suited to win MT benchmark runs against the eventual 8/24c intel competition. If the scheduler kinks are worked out, it should be the fastest in pretty much everything (except perhaps some high-clock ST scenarios)

 

[exquisitechar]

Then again not a fan of mobile chip replacing desktop. If i would be looking to get mobile level of performance, then i would be buying mobile device, not desktop.

It'll be better than how desktop is now. I'm looking forward to it.

 

[Timmah!]

Is Strix-Halo not enough CPU power?

Will its speed be infinite, so it will compute everything instantly?
More CPU power is always appreciated. Desktop trades performance for portability. If i am getting chip meant for portable devices, i am not taking one of 2 main advantages desktop is giving me (the other one being modular design).
Additionally, Strix-Halo seems to have bigger GPU. Thats waste on desktop with discreet graphics. I for one already have 2 4090s. I for one dont need crappy RDNA graphics in CPU on top of that. Use that die space for something else, like more CPU cores, for example.

Assuming both will cost around 999$, if you are a user of those applications and given the choice between 16 core Z5 vs 32 core Z5, which one would you choose? Would a 30% increase in single thread performance more important than near double performance for multi threaded applications?

This really depends on your particular workload. I use both apps, CAD for drawing, MAX for modeling. These are AFAIK single-threaded workloads. What is multi-threaded, is the rendering process. But i dont do that with Max, or better said CPU renderer, that would use those additional cores, since i use GPU based Octane for that purpose.

That said, its not like they could not do 32C chip, that would clock exactly the same as 16C in low threaded scenarios with only one of the CCDs being in use, and then clock lower to say 4,5GHz or whatever would be needed to fit the TDP when running MT tasks. And have best of both worlds.

 

[MadRat]

Instead of a simple focus on core count the future is focused on modular layouts that spread thermals on the micro, and optimizes I/O across the macro. Do you really want to cram cores onto one side of a chip, where the maximum load in one core impacts performance across all of the cores? Or would you rather spread them out and equalize thermals on maximum loads, smoothing out cooling requirements? Its not even all about heat but also proving clean power to each independent core group.

 

[Joe NYC]

Venice will have a new socket SP7

https://twitter.com/x/status/1700146795899162699

 

[Joe NYC]

Will its speed be infinite, so it will compute everything instantly?
More CPU power is always appreciated. Desktop trades performance for portability. If i am getting chip meant for portable devices, i am not taking one of 2 main advantages desktop is giving me (the other one being modular design).
Additionally, Strix-Halo seems to have bigger GPU. Thats waste on desktop with discreet graphics. I for one already have 2 4090s. I for one dont need crappy RDNA graphics in CPU on top of that. Use that die space for something else, like more CPU cores, for example.

Maybe it just means the new architecture, dumping IFoP in favor in favor of of Fan Out packaging similar to RDNA3, which will have high bandwidth links. But in general,, quite inexpensive packaging.

But the Zen 5 CCDs are still shown as independent dies in MLID video. so that's what may be shared. We will see if there will be different IO Dies between high end Notebook, with more powerful iGPU and just basic one for desktop.

But Intel is going with stronger iGPU across the future product lines of MTL and ARL, from notebook to desktop.

 

[Joe NYC]

Same poster also posted across the vendors partial image of roadmap with flag of sensitive image in the tweet. Uploading image from the tweet:

If we take Falcon Shores as 2025, it would imply that Venice and Mi400 are also 2025

 

[Fjodor2001]

Apple LITTLE is an actual sub-600mW LITTLE.
AMD dense cores are not that.

No.
Poverty cores make sense in cost-constrained segments.
Desktop isn't one of them (and isn't really a thing Zen6 onwards).

You're missing the point. b.L will always perform better in MT scenarios if correctly designed, given the same technical constraints (power consumption, process tech node, etc). Both on mobile and on desktop. And yes both mobile and desktop are power constrained.

Trying to get more MT performance by keeping core count constant and instead increasing CPU frequency or creating very big complex cores is just so much less efficient. Those last few 100 MHz:es will increase the power consumption a lot, since power consumption increases exponentially for the highest frequencies. That extra power consumption could instead be spent by having additional smaller cores running at a bit lower frequency, thereby consuming much less power for the same MT performance increase. I.e. essentially better perf/watt.

The only reason to only have big cores is if you only care about ST performance. But it's very rare to need more than ~8 cores with max ST performance. For workloads requiring more than ~8 cores you usually want max MT performance at lowest possible power consumption, and for that b.L is a better architecture.

More or less everyone except AMD is already doing it. Qualcomm, Intel, Apple, etc. Yes for the desktop lines too, e.g. Mac Pro (Apple M2 Ultra chip has 16P/8E cores).

AMD is just a little late to the party. But they'll get there too eventually. Luckily they have good CPUs in the current Zen lines anyway, so they'll get by fine with those for the time being. It's just that they'll be doing even better when they also go b.L.

 

[adroc_thurston]

desktop are power constrained.

Just no, DT power went waaaay up over the past 20 years.

Trying to get more MT performance by keeping core count constant and instead increasing CPU frequency or creating very big complex cores is just so much less efficient

More IPC.
moooooooooore. more.

The only reason to only have big cores is if you only care about ST performance.

Uniform perf profile is worth a lot.
A lot more than you think, really.
Real workloads aren't cinememe.

More or less everyone except AMD is already doing it

DT game is just Intel and AMD.
Everyone else works in cost-constrained markets where poverty cores are necessary.

Yes for the desktop lines too, e.g. Mac Pro (Apple M2 Ultra chip has 16P/8E cores).

Apple LITTLE is an actual LITTLE, made for background work and background work alone, in phones, tablets or laptops.

You want AMD to sell you more cores for cheaper, but that just ain't happening for a looooooong while outside of laptop.
Sorry, you'll have to make do with a 16c part or opt for a Threadripper, and either way is a margin win for AMD.

 

[Fjodor2001]

You got your personal priorities mixed up with mainstream consumer priorities. For the consumer base, core count beyond 16 is not even a remote priority. Today consumers need faster ST and lower cost while hitting 8+ core count in the process.

Anyone who is seriously in need for more than 16 fast cores is a pro-sumer or professional, and their needs are not really representative of the mainstream consumer market. The only reason we see opinions like yours pop up every now and then is because both Intel and AMD have neglected the HEDT segment, either through attrition or prohibitive pricing. However, feeling unrepresented in their product offering does not make you an exponent for mainstream consumer needs.

If you want to go into that kinda reasoning, how many mainstream consumers need more performance than 7950X provides, regardless of whether it's ST or MT performance?

 

[adroc_thurston]

how many mainstream consumers need more performance than 7950X provides

Mainstream consumers buy laptops.
Laptops need infinite amount of more 1T perf to make day to day javascript piles run faster.

 

[Fjodor2001]

Just no, DT power went waaaay up over the past 20 years.

We're already at 170W TDP. Have you not noticed how consumers are complaining about that?

You want to bump it to how much now, 300W? And for what reason, when you can achieve the same MT performance at lower power consumption by using b.L instead?

More IPC.

It's not either or. You can do b.L while still getting better IPC. Look at Apple for example.

DT game is just Intel and AMD.
Everyone else works in cost-constrained markets where poverty cores are necessary.

So why do you think Apple is using b.L on DT? They are certainly not cost constrained.

Apple LITTLE is an actual LITTLE, made for background work and background work alone, in phones, tablets or laptops.

Apple have b.L in their desktop/workstation CPUs too. The Mac Pro (do not mix up with Mac Book Pro which is a laptop!) as I mentioned uses Apple M2 Ultra chip which has 16P/8E cores.

You want AMD to sell you more cores for cheaper, but that just ain't happening for a looooooong while outside of laptop.
Sorry, you'll have to make do with a 16c part or opt for a Threadripper, and either way is a margin win for AMD.

I want whatever CPU fits my needs at the lowest possible cost. Same as everyone else.

I don't care if it's Intel, AMD, or something else. If AMD does not increase MT performance sufficiently (which most likely requires core count increase), I might have to go with Intel instead in the future to get better MT performance. And if none of them provides it at a reasonable cost, I won't upgrade. Either way, not sure how that is a win for AMD.

 

[Fjodor2001]

Mainstream consumers buy laptops.

So why have the discussion about what mainstream consumers want on desktop at all w.r.t. ST/MT performance then, if we're assuming only power users buy desktop PCs anyway? And it was not me who initiated that discussion BTW.

Laptops need infinite amount of more 1T perf to make day to day javascript piles run faster.

Why are they using b.L on laptops then? Even AMD is doing it (soon - with Zen5).

 

[adroc_thurston]

Have you not noticed how consumers are complaining about that?

No.

Look at Apple for example.

Apple loads on more big cores across the stack, keeping the LITTLE cluster exactly the same everywhere (it's 4C + what, 3MB L2?).

The Mac Pro (do not mix up with Mac Book Pro which is a laptop!) as I mentioned uses Apple M2 Ultra chip which has 16P/8E cores.

It's a CoWoS'd laptop chip and those LITTLEs aren't for nT spam at all.

If AMD does not increase MT performance sufficiently

They don't need to, since 1T is the primary selling point everywhere.

So why have the discussion about what mainstream consumers want on desktop at all w.r.t. ST/MT performance then, if we're assuming only power users buy desktop PCs anyway?

Because most people doing DYI PCs do it for gaming, and gaming perf doesn't require 16c parts.

Why are they using b.L on laptops then?

Poverty segment where costs matter.

 

[Glo.]

So why do you think Apple is using b.L on DT? They are certainly not cost constrained.

Because they have to scale up their mobile chips, not scale down the desktop chips.

And yes, part of it is to save money.