Question Zen 6 Speculation Thread

[Josh128]

Well even if not official the leaks clearly point to 10% IPC as a baseline goal, Z4 did 13-14%, Z2 did 15%. To hit +70% socket perf you are going to need perhaps a little more than 10% unless the new memory/interconnect/uncore carry hard, or the clocks are utterly insane.
ctively better.

The nT clocks will be utterly insane compared to Zen 5. The uplift will be at least equal, but possibly more, than the +20% nT clocks Zen 4 brought over Zen 3. That is basically a given.

1T/fmax clock is the question. Will it hit the fabled ~6.5GHz or not? At +10% IPC, it will need 6.5GHz to attain +25% 1T perf over Zen 5 in Cinememe R23. 25%-29% is the average 1T uplift Zen 4 got over Zen 3, so Im sure they are doing whatever they can to at least land in the >20% area. I hope they make it, but that kind of speed is completely unprecedented, thats why Im still skeptical they will be able to do it.

 

[adroc_thurston]

Will it hit the fabled ~6.5GHz or not? At +10% IPC, it will need 6.5GHz to attain +25% 1T perf over Zen 5 in Cinememe R23

Cinememe IPC will be like 2%, who cares.

but that kind of speed is completely unprecedented

It's 14% fmax spread over 2 shrinks.

 

[Joe NYC]

?

Using Q1 2025 (because Q2 has some one time elements)

Q1 2025 Operating Income
Intel CCG only: $2.4 billion
AMD entire company: $1.3 billion

I find it dumb to scoff at Intel "low margin" client division. It still is a freaking golden nugget.

Yeah they did, that's why LNL turned from a $1200 laptop part into a $700 one.

The direction was to further margin contraction. And Intel said it is having hard time keeping up with the demand of the cheapest on of all - Raptor Lake.

Then this news showed up:
(maybe reversal of strategy by LBT?)

https://twitter.com/x/status/1971028333359137178

 

[adroc_thurston]

I find it dumb to scoff at Intel "low margin" client division. It still is a freaking golden nugget

Yeah it is, costs are ramping quickly while ASPs are stangant.

The direction was to further margin contraction

Wha-

 

[Joe NYC]

Based on a little reading and research, AMX is only interesting in a few use cases. It is severely outclassed (by many orders of magnitude) by accelerator cards from AMD and NVidia.

I am not sure where that leaves AMX.

Do you have a specific use case where it will benefit DC applications (Not LLM which will be performed on accelerators for sure).

It seems to me that precisely because the gap between CPU only and CPU+GPU is so wide, and cost of adding the GPU is also high and has a lot of extra complexities, it seems that AMX may have found a niche in this gap.

 

[Joe NYC]

Yeah it is, costs are ramping quickly while ASPs are stangant.

That is true, as Intel shifts from Raptor Lake to chiplets and Foveros Tax.

Wha-

It seems that LBT wants to raise prices and ASPs. We will see how it works out.

Maybe it will work out near term, because demand in H2 seems to be strong. Longer term: TBD

 

[Josh128]

Using Q1 2025 (because Q2 has some one time elements)

Q1 2025 Operating Income
Intel CCG only: $2.4 billion
AMD entire company: $1.3 billion

I find it dumb to scoff at Intel "low margin" client division. It still is a freaking golden nugget.



The direction was to further margin contraction. And Intel said it is having hard time keeping up with the demand of the cheapest on of all - Raptor Lake.

Then this news showed up:
(maybe reversal of strategy by LBT?)

https://twitter.com/x/status/1971028333359137178

Seems that would only push sensible people to AMD, if true.

 

[Josh128]

Cinememe IPC will be like 2%, who cares.

I certainly hope its more than that, but you might be right. It'd be nice to finally snatch that "trophy" from Intel, who've held it since Raptor Lake release (for x86 at least).

 

[Joe NYC]

Based on a little reading and research, AMX is only interesting in a few use cases. It is severely outclassed (by many orders of magnitude) by accelerator cards from AMD and NVidia.

I am not sure where that leaves AMX.

Do you have a specific use case where it will benefit DC applications (Not LLM which will be performed on accelerators for sure).

Just to add to my previous post: Suppose it is true that this is a niche and it benefits only a few use cases.

But it just happens that this niche is ballooning into 1/2 trillion business. What may have been a valid answer 2 years ago may no longer be a valid answer now if the market has grown by factor of 100x.

 

[Joe NYC]

Which would be sad, if it took them two years to ship.

How long was the hype of the previous version before product actually shipped? It seems like well over a year.

 

[Joe NYC]

If it is gonna be N2P, it can't be released anytime in 2026 for the same reason nothing N2 is getting released in 2025.

We were discussing one of the dies that is N3P (Medusa Point) that does not need to wait for N2P - why this die does not ship earlier in 2026, such as Mid 2026 or even H1 2026.

 

[DrMrLordX]

We were discussing one of the dies that is N3P (Medusa Point) that does not need to wait for N2P - why this die does not ship earlier in 2026, such as Mid 2026 or even H1 2026.

The design might not be finalized by that point for any Medusa variant.

 

[adroc_thurston]

The design might not be finalized by that point for any Medusa variant.

?

 

[branch_suggestion]

What is exactly stopping the CCD reuse laptop parts from having similar 1T boost clocks to desktop?
Assuming 1T power is comfortably below TDP.

 

[adroc_thurston]

What is exactly stopping the CCD reuse laptop parts from having similar 1T boost clocks to desktop?

Nothing.

Assuming 1T power is comfortably below TDP.

N2 limits the Vmax you can dial in safely anyway.

 

[reaperrr3]

Hmm. It seems some rumor monger managed to re-define expectations. Well, not problem.

For the record, I was mostly talking desktop vs. desktop, so 5.7 -> 6.3-6.5.

If Zen4 could up the max turbo by 800 mhz from a very mature N7 to N5, why not Zen6 with a jump from N4P to N2P (and allegedly later N2X)?

 

[poke01]

For the record, I was mostly talking desktop vs. desktop, so 5.7 -> 6.3-6.5.

If Zen4 could up the max turbo by 800 mhz from a very mature N7 to N5, why not Zen6 with a jump from N4P to N2P (and allegedly later N2X)?

I think it’s possible. Turin-D got a massive speed boost

 

[adroc_thurston]

Turin-D got a massive speed boost

So does Venice-D.

 

[511]

I was only thinking of Zen6 at 6.3-6.4 Ghz

 

[adroc_thurston]

I was only thinking of Zen6 at 6.3-6.4 Ghz

that's 12% fmax. also normal.
It'll be in the range basically, think 12-15% fmax depending on the part.

 

[511]

Someone will dethrone 6.2 GHz RPL finally 🤣🤣.

 

[adroc_thurston]

Someone will dethrone 6.2 GHz RPL finally 🤣🤣.

hey this time without the boom-boom

 

[511]

hey this time without the boom-boom

It's seppuku though not boom boom

 

[OneEng2]

AMD's opinion on AMX is very clear, just buy GPUs... They even had a TCO chart to that effect...

... and considering workloads that AMX appears to be targeted at, I have to agree with AMD on this one.

and that the performance increase from the 9965 to the top of SKU Venice CPU is ~1.7x...

Yea, but core count increase (192 -> 256) is doing lots of that. Still, I suspect that N2 will allow higher clocks, and the new socket allows for higher power. Finally, higher memory per core frees up some performance as well.

I don't think much of it is because of IPC.

That's just harder to do when you can't triple your 1t power in 2 years.

It's flat out impossible IMO.

If it is gonna be N2P, it can't be released anytime in 2026 for the same reason nothing N2 is getting released in 2025.

I agree. I'm not sure where everyone's timing is coming from.

It's Zen6 with a new (single) IOD instead of the dual-IOD of Venice.

That makes a great deal more sense than thinking Zen 7 would be released so soon (which makes no sense at all).

It's great that AMD is going to use it to maximum advantage in server, in Zen 7, but client CPU market is 2x the size of server CPU market.

Not last time I checked. Link please. Please use revenue figures, not unit sales.

But ARM vendors are increasing frequency faster than AMD. Go plot a graph from 2022 if you need to.

ARM is at such low clocks that it is much easier to raise frequency .... especially if you also raise power. This gets much more difficult as you move forward (exponentially so).

Note that Zen3 didn't look that impressive outside of games and CCD-latency-sensitive workloads, either.

Zen 5 to Zen 6 will be more like Zen 4 to Zen 5 (my guess).

I find it dumb to scoff at Intel "low margin" client division. It still is a freaking golden nugget.

Agree that it is dumb to scoff. Not sure how golden it is though. I have to wonder how much of the profit comes from high end stuff vs OEM mid range and below.

 

[adroc_thurston]

It's seppuku though not boom boom

wdym seppuku, 14900k(s) was just doing some good honest jihad with a wee little bit of suicide carbombing.