Question Zen 6 Speculation Thread

[adroc_thurston]

So the CCDs are identical this time between server, desktop and desktop replacement?

not really? They're still different Vt's et al.

Let's just say I'll believe it once I see it for myself.

that's just how 2.5D slabs work.

 

[OneEng2]

Can't believe this needs to be stated, but why on earth are you comparing CCD area (meaning, only cores + IFOP) and the whole SoC, that has the CPU cores, the NPU, memory controllers and a relatively big iGPU? This is not an apples-to-apples comparison.

To the best of my understanding, the whole SoC is quite large. A single CCD is quite small. Don't look at it as the cumulation of CCD's is X mm2. Look at it as "I can get X number of good products from Y number of wafers" and "Total cost of one big SoC is X and Total cost of chiplets to do the same thing is Y".

Chiplets have proven to be a superior method of achieving the most value per $. Monolithic designs have proven to be superior in overall performance.

Chiplets also offer design modularity at a level beyond what can be reached with monolithic designs.

So my point is that should AMD desire to match the performance of M5 with Zen 6, and cost wasn't a factor, and modularity wasn't a factor, and scalability wasn't a factor, they could do it, but the design would be quite different than the Zen 6 we will be getting.

So it is an Apples to Apples comparison. Zen 6 is a more saleable design in both number of cores AND performance per $ (and likely design effort efficiency across a broad range of products).

If AMD's Zen 6 design targets were the same as M4, I am certain Zen 6 would look much more like M4.

 

[poke01]

To the best of my understanding, the whole SoC is quite large. A single CCD is quite small. Don't look at it as the cumulation of CCD's is X mm2. Look at it as "I can get X number of good products from Y number of wafers" and "Total cost of one big SoC is X and Total cost of chiplets to do the same thing is Y".

Chiplets have proven to be a superior method of achieving the most value per $. Monolithic designs have proven to be superior in overall performance.

Chiplets also offer design modularity at a level beyond what can be reached with monolithic designs.

So my point is that should AMD desire to match the performance of M5 with Zen 6, and cost wasn't a factor, and modularity wasn't a factor, and scalability wasn't a factor, they could do it, but the design would be quite different than the Zen 6 we will be getting.

So it is an Apples to Apples comparison. Zen 6 is a more saleable design in both number of cores AND performance per $ (and likely design effort efficiency across a broad range of products).

If AMD's Zen 6 design targets were the same as M4, I am certain Zen 6 would look much more like M4.

You do know the most comparable product to M4 that AMD makes is Strix Point, a monolithic laptop SoC?

 

[adroc_thurston]

You do know the most comparable product to M4 that AMD makes is Strix Point, a monolithic laptop SoC?

AMD doesn't make tablet parts.

 

[poke01]

AMD doesn't make tablet parts.

The point I’m trying to make is that AMD also makes monolithic SoCs for mobile

 

[OneEng2]

You do know the most comparable product to M4 that AMD makes is Strix Point, a monolithic laptop SoC?

Good point.

232.5 mm2; however, its performance is not as good as Strix Halo which is a chiplet design.

Still, if you look at the performance per $, I wonder how this pans out?

 

[511]

M4 is 168mm2 N3E and STXP is 232mm2 N4P there is a huge density gap

 

[OneEng2]

M4 is 168mm2 N3E and STXP is 232mm2 N4P there is a huge density gap

15%?

 

[511]

15%?

Density as in node density gap N4P vs N3P and a full node shrink

 

[marees]

No zen 6 client desktop on 'N2X' ?

https://twitter.com/x/status/1981901613636985202

 

[Tigerick]

Geez, sometimes the roadmap is so stupid and yet people still believe Zen6 12-core CCD will be fabbed by N2P/N2X....



Let me repeat: Venice is 32-core CCD fabbed by N2; not N2P or N2X. Meanwhile 12-core Zen6 will be fabbed by N3P/N3X. I am more inclined to believe it's N3X, same as Zen5 CCD, N4X. The rest of Zen 6 APUs and Soundwave will be fabbed by N3P. And it is aligned with TSMC's roadmap. That' why it is funny MLID speculated Zen6 will be delayed based on AMD roadmap; even though TSMC roadmap shown something more believable.

 

[Joe NYC]

No zen 6 client desktop on 'N2X' ?

https://twitter.com/x/status/1981901613636985202

Far more likely that Zen 6, 12 core chiplet for desktop will be the same as for server and mobile (N2P) than delayed to mid 2027 waiting for N2X.

 

[yuri69]

Let me repeat: Venice is 32-core CCD fabbed by N2; not N2P or N2X. Meanwhile 12-core Zen6 will be fabbed by N3P/N3X.

Yup, this sounds like the classic AMD/cheapo way.

 

[ETI4711]

ワークステーション/PCで存在感を増すAMD、CVP＆GMのMcAfee氏に聞いた現在と未来

AMDのDavid McAfee氏が8月21日に来日し、30分程ではあるが時間を戴いて色々お話することが出来た。今回はプレゼンテーションも何もなく、いきなりQ＆Aの形式でのスタートであったが、その模様をお届けしたい。

news.mynavi.jp

Q: By the way, CEO Lisa Su has already introduced the core of the Zen 6-based EPYC product manufactured on TSMC N2 (2nm process) as "Venice" (development code name). Why Venice?

David McAfee: What do you mean?

Q: Venice was the code name for the Athlon 64 with 512KB L2 and manufactured on a 90nm SOI process. You joined AMD in 2006, when Venice-based Athlon 64s were still on the market, right?

David McAfee: I see. We transitioned the original Venice to TSMC's 2nm (laughs).

Seriously, as you know, we use the same processors for our servers and desktops, so we expect the same architecture that Lisa showed with Venice to be available on the desktop.

---
And btw. AMD is scheduled to deliver Helios to Oracle in Q3 2026. Venice is expected to be integrated into Helios.

 

[luro]

No zen 6 client desktop on 'N2X' ?

https://twitter.com/x/status/1981901613636985202

It never was on N2X

 

[basix]

N2P is a big jump from N4P. If they reach clock targets with N2P, why bother with N2X for +5% clock rate and risk delays? Especially when the sole reason should be CCDs to reach high desktop clock rates (smallest of all markets). Server does not need it. Mobile does not need it. Consoles don't need it.

And I assume the "N2P" variant AMD is using is still somewhat customized for Zen 6.

 

[ETI4711]

N2P is a big jump from N4P. If they reach clock targets with N2P, why bother with N2X for +5% clock rate and risk delays?

From the TSMC Symposium 2025:

https://www.tsmc.com/english/node/223# Technology Leadership Dr. Yuh-Jier Mii

N2X is for production in 2027, too late for Zen 6.
N2P is on track for production in 2nd half of 2026. This is start HVM, not first products with N2P delivered.

The numbers are N2P vs N3E. Why do you think TSMC is comparing not to N2?

 

[basix]

I assume N2X gets compared to N2 vanilla on that slide. And as N2P already brings a Fmax boost, N2X won't bring too huge gains.

But anyways, the point is that the jump from N4P to N2P is much bigger than N2P to N2X. Maybe 80% of the overall Fmax gains TSMC states. So N2X might be good for Zen 6 desktop but it will not win the game.

 

[adroc_thurston]

you guys are reading so hard into vaguepoasted foundry roadmap that has nothing to do with how AMD/TSM relationship works.

 

[Meteor Late]

N2P is a big jump from N4P. If they reach clock targets with N2P, why bother with N2X for +5% clock rate and risk delays?

What do you mean why? of course, to jerk off with the 7GHz pipedream!

 

[Meteor Late]

you guys are reading so hard into vaguepoasted foundry roadmap that has nothing to do with how AMD/TSM relationship works.

So will AMD use N2X for desktop Zen6 or not?

 

[adroc_thurston]

So will AMD use N2X for desktop Zen6 or not?

They'll use whatever DTCO's mess they opt for.
Will they use overdrive xtors with zoomzoom Vt's? probably.

 

[Josh128]

Let me repeat: Venice is 32-core CCD fabbed by N2; not N2P or N2X. Meanwhile 12-core Zen6 will be fabbed by N3P/N3X.

This is what Ive been postulating for a long time now and have taken insults for even suggesting it.

 

[adroc_thurston]

This is what Ive been postulating for a long time now and have taken insults for even suggesting it.

yeah cuz it taped out a while ago on N2p.
You do get N3p, in mainstream mobile.

 

[reaperrr3]

yeah cuz it taped out a while ago on N2p.

Can we even consider N2P/X separate processes at all?
As far as I understood, one of the newer N3 chips (forgot if it was Snapdragon X2 or some Nvidia chip) uses N3P with some N3X transistors for clock-critical parts, so I'd assume this kind of mix is possible with N2P/X too and AMD would only design one 12C CCD, and not one N2P variant and one N2X variant.