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

[DisEnchantment]

Speculate at will

 

[DrMrLordX]

There is no desktop Zen6, it's just client.

Sounds ominous, actually.

 

[adroc_thurston]

Sounds ominous, actually.

Kinda the opposite, DT won't be stuck with unfitting uncore anymore.

 

[DrMrLordX]

Kinda the opposite, DT won't be stuck with unfitting uncore anymore.

Ehh

In theory desktop could benefit from sharing a design with mobile. In practice, desktop just gets shuffled further back on the priority list.

 

[FlameTail]

Zen

 

[adroc_thurston]

desktop just gets shuffled further back on the priority list.

Desktop is also luggables, ergo mobile, ergo relevant.

In theory desktop could benefit from sharing a design with mobile

It's a good fit.

 

[biostud]

I understand your pain. X670E is way, way overpriced for what it offers, especially when there are $150 mobos from Gigabyte/ASROCK that can run a 7950X without much issue.

You can get X670 boards from under $200 and X670E from $230

 

[EXCellR8]

man i'd kill for a new chipset with boards that weren't all the same gray/black meh color scheme. i'm not saying we go back to green and yellow PCI slots but someone give these modern boards some character! it's like a sea of monochrome...

/uselessrant

 

[Joe NYC]

There is no desktop Zen6, it's just client.

Seems like MLID does not have much info on Zen 6 client so far...

Server parts live on their own and there should be major differences between mainstream (SP8) and halo (SP7) platforms.

It seems like the dense 32 core die is now the mainstream version. Given how well the Zen 4c performs on the server workloads, it's not surprising.

So if the Siena / Sorano / Next becomes the mainstream part and it goes up to 2 I/O dies, then, there would be either 4 ports for CCDs available, so up to 128 cores. Or up to 6 ports if something can be attached to the sides of the IO dies.

Are these connections between IO Dies and CCDs using Active Silicon Bridges?

 

[Joe NYC]

Damn it, I think I know why there is no Zen7 in the roadmap. Cause 32-core Venice CCD made by N2 process will become Zen7 which will be unveiled in 2027. By that time, DDR6 is readily available, AMD only need to redesign new IOD with 128-bit DDR6 memory bus then voila we have new platform for AM6..

I think the IO die would be an overkill for mainstream desktop.

But it would be interesting if AMD started an HEDT platform using a single IO die, using a socket bigger than AM4.

The problem with Threadripper is that it is not that much different from the server parts, and the cost is near the server parts.

What HEDT platform needs is something that is about half way between desktop and server, or closer to desktop than the server. One I/O die might be the right spec.

The only thing is if there would be enough demand. There should be more demand for it than for Threadripper, which really has just miniscule demand...

 

[igor_kavinski]

man i'd kill for a new chipset with boards that weren't all the same gray/black meh color scheme.

https://www.newegg.com/gigabyte-x670e-aorus-pro-x/p/N82E16813145485

Out of stock, though.

 

[StefanR5R]

…what I for one understood from this thread so far…

I’ve not followed all posts in this thread in detail, but cannot recall that such a definitive conclusion has been made.

I spoke just for myself. :-)

Regardless, isn’t all the discussion in this thread just speculation anyway? I’ve not seen any official info from AMD mentioning what you described.

Correct, apart from AMD claiming that Zen 5 was a "new grounds-up microarchitecture" with "re-pipelined front end and wide issue". And AMD spoke of Zen 5 products to be made on 4nm and 3nm class manufacturing nodes (and further, that there shall be Zen 5/ Zen 5c/ Zen 5 V-cache products not entirely unlike the current product split).

When did AMD last post updates of their CPU roadmaps anyway? Seems as if there weren't any since their Financial Analysts Day in June 2022.

Beyond what AMD have been publishing, there seems to be a consensus here that (at least with those products which will be made on 4nm), the advancements of the manufacturing node(s) used for Zen 5 won't be huge compared to those used for Zen 4. This provides a certain bracket for the all-core performance changes that can be expected. All-core performance depends a lot on (a) power budget and (b) power efficiency. There are more factors obviously, but these two stick out. Regarding (b) power efficiency, large jumps won't be made with microarchitectural changes alone, empirically.

As for (a) the power budget's influence on all-core performance — and that's just my understanding, not necessarily a consensus and definitely nothing official — : Those CPU SKUs of the Zen 4 generation which are configured to work very far away from the V/f sweet spot might be succeeded by corresponding SKUs of the Zen 5 generation which are driven somewhat nearer to that sweet spot. In such cases, a noteworthy upgrade to all-core throughput could be had even at same power limit. Conversely, if we take SKUs of the Zen4 product range which run at quite efficient points of the V/f curve under all-core loads, then their direct successor SKUs can be expected to provide only a minor step up in all-core throughput at same power limit, but still might profit from their widened cores if given an increased power budget. Am I talking nonsense, or am I merely channeling Captain Obvious here?

 

[igor_kavinski]

Am I talking nonsense, or am I merely channeling Captain Obvious here?

I think that Zen 5, being a more efficient arch in terms of IPC, will get hotter than Zen 4 because it will be less busy between pipeline stalls and loading data from caches/RAM. That, to me, explains the rumored lower maximum frequencies. So while Zen 5 may be able to consume more power and scale to higher performance, it won't be able to do so without exotic cooling.

 

[StefanR5R]

Wider core architecture on a somewhat similar area size = denser logic could correlate with higher heat flux density at same clock, but it also could mean that, compared to the Zen 4 (non-c) speed daemon, less of the area can be spent on circuitry that's dedicated to the purpose of reaching high clock speeds.

 

[adroc_thurston]

It seems like the dense 32 core die is now the mainstream version

Absolutely not.

 

[Timmah!]

Desktop is also luggables, ergo mobile, ergo relevant.

It's a good fit.

Nah, i am buying desktop, cause i dont need mobility. And i expect significantly more performance in return. Currently, the mobile 7945hx provides like 90 percent of performance of desktop 7950x, both in ST and MT. As if that was already meh not enough, conflating desktop and mobile into client or whatever you call it, takes this another step further. If i specifically did not need those top Nvidia gpus for gpu rendering, i might as well buy laptop as my next rig.

 

[adroc_thurston]

Nah, i am buying desktop, cause i dont need mobilit

You also don't need server uncore so with Zen6 you're getting a more appropriate solution.
You're welcome.

 

[Timmah!]

You also don't need server uncore so with Zen6 you're getting a more appropriate solution.
You're welcome.

Uncore is the IOD die?
Anyway, if getting something i dont need alongside something i need, is the only way, then fine, gimme both.

 

[igor_kavinski]

You also don't need server uncore so with Zen6 you're getting a more appropriate solution.

So with Zen 6, we will finally have a dedicated client solution and not rejected server silicon?

 

[adroc_thurston]

Uncore is the IOD die?

Along with fabric topology and d2d link choices.

So with Zen 6, we will finally have a dedicated client solution and not rejected server silicon?

Well yeah.

 

[igor_kavinski]

Well yeah.

What will be the fate of the rejected server silicon once Zen 6 ships?

 

[Ajay]

Ehh

In theory desktop could benefit from sharing a design with mobile. In practice, desktop just gets shuffled further back on the priority list.

I'm curious to see this. As I've been building DT DIY systems for 25 years, it, uh, feels weird. But, if it performs well, IDC.

 

[igor_kavinski]

In theory desktop could benefit from sharing a design with mobile.

Benefit how? Desktop will get gimped if it's based on mobile silicon. It would become the same thing as an HX series gamer/workstation CPU.

 

[H433x0n]

Now that I think about it - I wish they took this approach this generation.

Imagine a Phoenix die with vcache that you could drop into an AM5 board. A monolithic AMD processor on N4P with a much stronger iGPU would dethrone the 7800X3D.

 

[H433x0n]

Benefit how? Desktop will get gimped if it's based on mobile silicon. It would become the same thing as an HX series gamer/workstation CPU.

Desktop is already gimped from being server table scraps. It should really just be Epyc & Threadripper using those dies. They could create 2 dies for all of mobile and desktop, 8C & 16C Phoenix.

If they were using Phoenix dies instead they’d probably have a stronger performing product (compared to 7600X & 7700X) and they’d have more flexible inventory for mobile client sales.

 

[adroc_thurston]

What will be the fate of the rejected server silicon once Zen 6 ships?

Death.

It would become the same thing as an HX series gamer/workstation CPU.

The only difference between that and DT is power target.

A monolithic AMD processor on N4P

Nothing changes besides fabric power.