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

[DisEnchantment]

Speculate at will

 

[misuspita]

I hope to see them in minipcs. Better form factor to cool 125W+. I mean, with a proper cooler, they can dissipate easily 180W (like those Zotac mini PC's)

I assume Strix Halo will mainly be in MBP wannabe laptops. Solder should be okay.

 

[Glo.]

I hope to see them in minipcs. Better form factor to cool 125W+. I mean, with a proper cooler, they can dissipate easily 180W (like those Zotac mini PC's)

One big cooler, like Deepcool AK500 and you are set.

Benefits of unified memory architecture.

 

[Glo.]

What are the chances of Strix Halo being delayed to push the production and use of LPCAMM during this 2024? Because i just realised it could be all be connected. Having a super high end laptop with soldered memory could be an issue, and LPCAMM solves that, it would also open the possibility of desktop motherboards, altrought AMD itseft might the one that would not want that.

Everybody wants their products soldered. Higher manufacturing costs, true, but at the same time, higher premiums, because of atypical solution, that requires more atypical solutions.

Apple, anyone?

 

[adroc_thurston]

What are the chances of Strix Halo being delayed to push the production and use of LPCAMM during this 2024?

0.

Benefits of unified memory architecture.

The only real benefits per se are for optimized GPGPU stuff and nothing client.

 

[maddie]

Everybody wants their products soldered. Higher manufacturing costs, true, but at the same time, higher premiums, because of atypical solution, that requires more atypical solutions.

Apple, anyone?

Soldered is higher cost?

 

[Glo.]

Soldered is higher cost?

Mobo + SOC + RAM is lower cost, compared to just the socketed CPU, separate from the MoBo?

 

[HurleyBird]

0.

The only real benefits per se are for optimized GPGPU stuff and nothing client.

Well, that and cost/capacity.

 

[adroc_thurston]

Well, that and cost/capacity.

VRAM capacity isn't a big limitation for most gaming laptops it's whatever.
Cost, yes, a sizeable net win.
Let's see how it does in the market.

 

[HurleyBird]

VRAM capacity isn't a big limitation for most gaming laptops it's whatever.
Cost, yes, a sizeable net win.
Let's see how it does in the market.

A Mac Pro/Mini form factor with a reasonable GPU + mountains of memory would have an interesting niche.

 

[adroc_thurston]

A Mac Pro/Mini form factor with a reasonable GPU + mountains of memory would have an interesting niche.

Yeah very useful for ML once SPIR-V support gets mainlined in ROCm.

 

[DrMrLordX]

So Intel fills up TSMC fabs, then gets AMD as IFS customer?

TSMC generally doesn't work like that. If Intel tried to "fill up" their capacity, they'd do the same thing they did with Intel already: charge them a ton of money and build out more capacity. TSMC isn't about to leave Apple or AMD hanging out to dry.

 

[soresu]

TSMC isn't about to leave Apple or AMD hanging out to dry

Certainly not for a customer that is basically a competing fab operator too.

 

[Shivansps]

Mobo + SOC + RAM is lower cost, compared to just the socketed CPU, separate from the MoBo?

Soldered is definitely lower cost(at least for the mass produced low end, or you have 2-3 version max like apple), this is why you have notebooks with a soldered DDR4 and a SODIMM slot. The motherboard cost increases but the overall BOM for the whole product is lower.

But this is not 4-8GB of DDR4, here we are talking about soldering at 32-64 GB of LPDDR5X for the 40CU one, thats kind of a lot. Upgradeable ram allows them to launch products with 8, 12 and 16GB of LPDDR5X and allow clients to upgrade later.

Now on high end products im not so sure if soldered is cheaper, specially in this case, what will be cheaper? producing 1000 motherboards or a CPU SKU with LPCAMM memory, or producing 100 with 8GB, other 100 with 12GB, other 100 with 16gb, etc?... i dont know, there is also the problem of producing too much of one and too little of other, etc.

This also allows for desktop boards to be built, and could be usefull in minipcs, the tech is already there, they just need to wait for production.

 

[eek2121]

From what I understand, AMD is looking at using alternative fabs for some of the cheaper/lower margin stuff, but unless China starts a war, the premium stuff is staying at TSMC.

I have not seen any concrete plans, but I do know conversations have been had regarding both CPU and GPU designs.

 

[Fjodor2001]

Regarding Strix Halo (intended for laptop), rumors say it'll be 16C Zen5 + 40 RDNA 3.5-spec compute units. So essentially close to DT variants of Zen5 in CPU perf (but likely with somewhat slower clocks), and extremely much higher iGPU perf than DT. This is assuming Zen5 on DT will only be max 16C and have same IOD (and iGPU) as Zen4.

So this means for Zen5, DT will not have much higher MT CPU perf than laptop, and much lower iGPU perf than laptop? And all released close in time?

 

[adroc_thurston]

This is assuming Zen5 on DT will only be max 16C and have same IOD (and iGPU) as Zen4.

Yea.

So this means for Zen5, DT will not have much higher MT CPU perf than laptop, and much lower iGPU perf than laptop?

Yea. It also has better d2d links.
-halo is a very not mainstream part though.

And all released close in time?

I don't think April'24 and January'25 are that close but I digress.

 

[Jan Olšan]

It exists but not the way you envision it.

Do you mean that Zen 6 desktop (the X/X3D SKUs) will stop having their own IOD and instead AMD fully moves to the APU dies to make the whole desktop lineup?

Do they add 3D V-Cache do the client APUs, or does the client APU have infinity links that will make it possible to connect a server CCD + V-Cache to an APU base?

 

[adroc_thurston]

Do you mean that Zen 6 desktop (the X/X3D SKUs) will stop having their own IOD and instead AMD fully moves to the APU dies to make the whole desktop lineup?

Medusa is disagg and I don't have any idea how many tiles.
Probably 2 or 3.

Do they add 3D V-Cache do the client APUs, or does the client APU have infinity links that will make it possible to connect a serve CCD + V-Cache to an APU base?

Server is its very own thing with Zen6.

 

[Fjodor2001]

I don't think April'24 and January'25 are that close but I digress.

Strix Halo for laptop is rumored to be 2024H2, and you said DT Zen5 is April 2024. So that means roughly 6 months apart, if we assume middle of 2024H2 for Strix Halo.

Weird that Zen5 laptop will be close to DT in CPU perf and hugely better in iGPU perf so close in time. Makes laptop look very good, and DT crap. Normally DT should outperform laptop by quite a bit.

 

[Glo.]

Strix Halo for laptop is rumored to be 2024H2, and you said DT Zen5 is April 2024. So that means roughly 6 months apart, if we assume middle of 2024H2 for Strix Halo.

Weird that Zen5 laptop will be close to DT in CPU perf and hugely better in iGPU perf so close in time. Makes laptop look very good, and DT crap. Normally DT should outperform laptop by quite a bit.

Strix Halo is late 2024/early 2025.

 

[Fjodor2001]

Strix Halo is late 2024/early 2025.

Well even if 9 months apart like adroc says, still close in time. DT will look bad, laptop good.

 

[gdansk]

DT will look bad, laptop good.

Laptop is like 80% of the market. Laptop good is what AMD wants.

 

[Glo.]

Laptop 80% of the market. Laptop good is what AMD wants.

Get used to the reality that more mobile first designs will appear in DT form factors.

With everything it means.

 

[Saylick]

Zen 5 hype train is starting to leave the station, for better or worse. Lol

https://twitter.com/x/status/1749473720182419515

https://twitter.com/x/status/1749455220894036097

 

[Thibsie]

Danilo ? Lol.
He's not a source of anything. Useless