Discussion Beyond zen 6

[Jan Olšan]

AM6 is a bit far out and is an LPDDR6x platform.

Don't you dare to be wrong on this one. HECK YEAH!

 

[adroc_thurston]

Don't you dare to be wrong on this one. HECK YEAH!

This is technically worse for desktop tho.

 

[511]

It's Socamm hopefully

 

[CouncilorIrissa]

AM6 is a bit far out and is an LPDDR6x platform.

God I hope they will have got their crap together by then by reducing the LPDDR latency to something competent aka PTL level. Having almost 130ns of latency on STX is disgusting.

 

[511]

God I hope they will have got their crap together by then by reducing the LPDDR latency to something competent aka PTL level. Having almost 130ns of latency on STX is disgusting.

NVL is doing SoC and Cores on different Tiles so i expect few % latency regression

 

[CouncilorIrissa]

NVL is doing SoC and Cores on different Tiles so i expect few % latency regression

Pretty sure you're in the wrong thread.

 

[adroc_thurston]

God I hope they will have got their crap together by then by reducing the LPDDR latency to something competent aka PTL level. Having almost 130ns of latency on STX is disgusting.

client

Hahahahhahahahahahahahhaha

 

[511]

Pretty sure you're in the wrong thread.

lmfao you are correct

 

[adroc_thurston]

lmfao you are correct

It's still true doe.
NVL latency is a regression from PTL.

 

[CouncilorIrissa]

Hahahahhahahahahahahahhaha

Ye I know, no flying duck given about client, yadda-yadda.
But 130ns is just horrific, even STXH has lower latency (not by much).

 

[adroc_thurston]

even STXH has lower latency (not by much).

stxH memlat is 150ns range.
Again, it's client. A joke segment.

 

[CouncilorIrissa]

stxH memlat is 150ns range.
Again, it's client. A joke segment.

It's around low-to-mid 120s judging by this test.

 

[adroc_thurston]

View attachment 141903
It's around low-to-mid 120s judging by this test.

Ah. Okay yeah.
Thought it was 150ns-ish.

 

[Joe NYC]

No.

Seems like 3/4 of the video is about 72c on AM5, which is clearly BS.

 

[Joe NYC]

Forget about dense outside of server and 3Dcore is not Zen7 afaik.

There is a difference between a 2 die CCD which starts as a 3D stacked product (vs. optional) vs. 3D core, where the core itself is split between multiple stacked dies.

Some of the things MLID says about Zen 7 being 3D does not make that distinction clearly, and I suspect there will be a lot of confusion until someone establishes a clear terminology.

 

[adroc_thurston]

Seems like 3/4 of the video is about 72c on AM5, which is clearly BS.

Dude's an idiot and has to clickbait.

the core itself is split between multiple stacked dies.

Yeah it's an upsized core built around sub-micron f2f SoIC.

 

[HurleyBird]

Zen 7 consumer 72c 🤔🤔🧐🧐

No.

MLID is just kinda awful at theory crafting. I can maybe see a 16 + 36 "CPU Titan" configuration, even though the asymmetry would be awkward. 36 + 36 is just nuts outside of maybe some niche not-for-consumers application. 36 + 36 would both be more expensive, and it would lose against 16 + 36 in practically everything consumers care about.

 

[OneEng2]

Zen 7 consumer 72c 🤔🤔🧐🧐

So what I got out of this is that Zen 7:

• %15-%20 IPC uplift (8% just from cache changes)
• Supports 16c CCD's vs 12c (Zen 6) on desktop
• May have a 36c CCD for DC for a 288c CPU
• Will double the L2 to 2M (still quite a bit behind ARL and NVL)
• Compute tile on A14
• Still on AM5
• 2028 release

I'm guessing only a moderate clock speed increase (5%), and a 2029 release IF NVL delivers. Otherwise, 2030 release.

I completely believe the 288c DC CPU. The 72c desktop on AM5? Wow. That sure seems like someone has been tipping into the liqueur cabinet 🙂

I believe we will definitely get a 32c/64t desktop variant (which is already questionably more than needed). I believe there will be workstation desktops with higher core counts (Threadripper), but not on AM5.

 

[adroc_thurston]

I can maybe see a 16 + 36 "CPU Titan" configuration

a) nope
b) nope.

%15-%20 IPC uplift (8% just from cache changes)

>20% and cache is whatever.

May have a 36c CCD for DC for a 288c CPU

It is a 36c stack.

I'm guessing only a moderate clock speed increase (5%), and a 2029 release IF NVL delivers. Otherwise, 2030 release.

Again, Intel does not exist in AMD roadmaps.
They do not care about Intel. It's not a thing.

 

[Jan Olšan]

This is technically worse for desktop tho.

I don't care about that as long as it improves power usage. Latency tends to not matter that much outside games and games want gamecache TM anyway.

 

[adroc_thurston]

I don't care about that as long as it improves power usage

That's an ATX issue.

 

[dangerman1337]

Launch is Q3'28.

I wonder if Zen 7 X3D may come out first for DIY, A14 looks too pricey and X3D tops the the charts often more than not anyways...

no seriously, what is the argument of putting Grimlock Ridge non-X3D for Desktop first and successors? People want that X3D cache if they got the cash for it.

 

[adroc_thurston]

I wonder if Zen 7 X3D may come out first for DIY

Maybe.

A14 looks too pricey

No, N2 and A14 are both very fair value shrinks.
Especially A14 since that one has proper SRAM scaling! Smaller bitcell and all. sexo.

no seriously, what is the argument of putting Grimlock Ridge non-X3D for Desktop first and successors?

They get to double-dip and do two rounds of ASP hikes. bretty simple.
AMD's a business, a very evil one, not a charity. If they can squeeze you for money, they will.

 

[Jan Olšan]

That's an ATX issue.

I'm ready to simp for ATX12VO too

 

[branch_suggestion]

So what I got out of this is that Zen 7:

• %15-%20 IPC uplift (8% just from cache changes)

He clearly said 15-25% but you do you.
And really they're gonna be targeting 20% bare minimum, OG Z5 on N3B was kinda that but they had to roll back and neuter a bunch of core resources.

• Supports 16c CCD's vs 12c (Zen 6) on desktop

Yes, 2 if you are rich.

• May have a 36c CCD for DC for a 288c CPU

Will, and it is a C3D, core die stacked on L3 die.
Core die doesn't have L3 I think but does have the mesh network that the L3 below connects to.

• Will double the L2 to 2M (still quite a bit behind ARL and NVL)

2MB is plenty, ARM basically tops out there and M5 Super is 1MB.

• Compute tile on A14

Yessir, H2 2028, always 2 years.

• Still on AM5

Yes, but people here already said that, Z7 uses the OMR IOD iirc.

• 2028 release

Yes

I'm guessing only a moderate clock speed increase (5%), and a 2029 release IF NVL delivers.

Clocks are probably the sameish as Z6 to drive IPC and efficiency.
It is 2028, NVL sucks.

Otherwise, 2030 release.

Z8 says hello.

I completely believe the 288c DC CPU. The 72c desktop on AM5? Wow. That sure seems like someone has been tipping into the liqueur cabinet 🙂

Not possible, IOD has only 2 stops and the dense stack uses 2, so you could get 1x 36c for embedded EPYC.

I believe we will definitely get a 32c/64t desktop variant (which is already questionably more than needed).

Indeed, 12c is already more than actually needed unless you hate SMT.

I believe there will be workstation desktops with higher core counts (Threadripper), but not on AM5.

SP8 TR is a given, SP7 depends on many things.