Discussion Zen 5 Discussion (EPYC Turin and Strix Point/Granite Ridge - Ryzen 8000)

[DisEnchantment]

Well, since many folks already got their hands (or at least going to get) on Zen 4 CPUs , time to discuss about Zen 5 (Zen 4 already old news )

We already got roadmaps and key technologies like AIE








Some things we already knew

• Dr. Lisa Su and Forrest Norrod already mentioned at FAD 2022 on May 9th, during Q&A that Zen 5 will come in N3 and N4/5 variants so it will be on multiple nodes.
• Mark Papermaster highlighted that it will be a grounds up architecture, Also mentioned last para here
• Mike Clark mentioned that they started to work on Zen 5 already in 2018. This means Zen 5 by the time it launches would have been under conception and planning and development for much longer than the original Zen program

For a CPU architecture launching in early 2024 in the form of Strix Point for OEM notebook refresh, tape out should be happening in the next few months already.
Share your thoughts

"I just wanted to close my eyes, go to sleep, and then wake up and buy this thing. I want to be in the future, this thing is awesome and it's going be so great - I can't wait for it." - Mike Clark

 

[Abwx]

No they aren't, jfc.

And yet it is , 14% better perf at same power in INT and 19% for FP for a 2S configuration...

AMD EPYC Milan Review Part 2: Testing 8 to 64 Cores in a Production Platform

www.anandtech.com

So you are just throwing wild guesses without even checking for the numbers...

Yes it is, lol.

It is not, that s not the debate but ARM implemented 8 ALUs wich have limited functionalities, hence the number, while AMD use 4 versatile ALUs wich have as much functionalities as all 8 ALUs on the ARM design.

 

[adroc_thurston]

And yet it is , 14% better perf at same power in INT and 19% for FP for a 2S configuration...

At 25% more power (they're 225W Rome and 280W Milan sockets).

but ARM implemented 8 ALUs wich have limited functionalities, hence the number, while AMD use 4 versatile ALUs wich have as much functionalities as all 8 ALUs on the ARM design.

Simple ALU ops is what you'll be doing 99% of the time jfc.

 

[HurleyBird]

Not the issue, the thing just had negative PPW over Zen2 in silly nT loads.

Which isn't relevant for server IPC.

 

[Abwx]

At 25% more power (they're 225W Rome and 280W Milan sockets).

As pointed by AT they had to increase the I:O power to cope with the overhead, but cores themselves are more efficient, FPU being the most power hungry 19% better perf cant be achieved with that TDP uplift, it would require 55-60% more CPU power to achieve 19% better perf at same design.

Simple ALU ops is what you'll be doing 99% of the time jfc.

ALUs are the real brains of the CPU, either they do INT related ops or are completing the FP ops, FP arithmetic computation being done of course in the FPU but the related datas manipulations are still done by the ALUs.

ARM use simples ALUs that have limited functionalities while AMD s Zen use complex ALUs that can perform more differents ops.

Arithmetic logic unit - Wikipedia

en.wikipedia.org

 

[itsmydamnation]

As pointed by AT they had to increase the I:O power to cope with the overhead, but cores themselves are more efficient, FPU being the most power hungry 19% better perf cant be achieved with that TDP uplift, it would require 55-60% more CPU power to achieve 19% better perf at same design.



ALUs are the real brains of the CPU, either they do INT related ops or are completing the FP ops, FP arithmetic computation being done of course in the FPU but the related datas manipulations are still done by the ALUs.

ARM use simples ALUs that have limited functionalities while AMD s Zen use complex ALUs that can perform more differents ops.

Arithmetic logic unit - Wikipedia

en.wikipedia.org

what is this gibberish ?

like 6 int operations cover like 95% of integer code.

you know

add/subtract , load data , store data , compare , branch , multiply

All modern CPU's execute uop's, everyone's are as simple or as complex as each other, ARM even have some "complex" ones x64 doesnt like those double loads.


I'm totally confused , i found Zen2 -> Zen3 spec int rate nT numbers matching what @adroc_thurston said. He then said AMD actually have nT in there hotchips '24 presentation for zne4 and instead of someone trying to find that we all start doing dumb things that involve not reading properly?

​

 

[adroc_thurston]

Which isn't relevant for server IPC.

Very relevant, power targets matter.

As pointed by AT they had to increase the I:O power to cope with the overhead,

Dawg it was a Daytona issue.

 

[HurleyBird]

Very relevant, power targets matter.

For clocks and performance, not IPC.

 

[Abwx]

what is this gibberish ?

like 6 int operations cover like 95% of integer code.

you know

add/subtract , load data , store data , compare , branch , multiply

All modern CPU's execute uop's, everyone's are as simple or as complex as each other, ARM even have some "complex" ones x64 doesnt like those double loads.


I'm totally confused , i found Zen2 -> Zen3 spec int rate nT numbers matching what @adroc_thurston said. He then said AMD actually have nT in there hotchips '24 presentation for zne4 and instead of someone trying to find that we all start doing dumb things that involve not reading properly?

​

I m aware that theres s 6 ops that are ubiquitous, but still, more ALUs seems necessary, otherwise BD s 2 ALUs wouldnt have been that much below Zen for IPC.

Beside i pointed that Zen 3 didnt have as much IPC uplift in MT than in ST, so you re just doing a false point here...

The numbers they provide in the Zen 5 slide for previous gen are these ones, Zen 3 didnt impove IPC for MT by 19% according to AT MT SPECint/fp tests.

AMD Zen 3 Ryzen Deep Dive Review: 5950X, 5900X, 5800X and 5600X Tested

www.anandtech.com

As you said things should be read properly otherwise answers do not make sense...

 

[mikk]

a) that's nT IPC (basically per-vCPU iso clk perf)
b) you're getting more!

7-10% (got 13%) and 8% (got 14%).

Late March or April.

March or April seems optimistic to me. AMD never specified the exact half year, they just said 2024. This is usually an indication for a H2 launch rather than a H1 launch. I don't think we are getting Zen 5 in H1 of 2024.

 

[adroc_thurston]

For clocks and performance, not IPC.

For IPC too, lol.
AMD has a pretty strict Cac ratio for each %% of IPC gained.

March or April seems optimistic to me

Right where it should be.

I don't think we are getting Zen 5 in H1 of 2024.

Everything not Strix-Halo is H1'24.

 

[itsmydamnation]

I m aware that theres s 6 ops that are ubiquitous, but still, more ALUs seems necessary, otherwise BD s 2 ALUs wouldnt have been that much below Zen for IPC.

Beside i pointed that Zen 3 didnt have as much IPC uplift in MT than in ST, so you re just doing a false point here...


As you said things should be read properly otherwise answers do not make sense...

SPEC INT RATE nT FOR MILIAN!!!!! FFS JESUS

 

[Abwx]

SPEC INT RATE nT FOR MILIAN!!!!! FFS JESUS

You should read through the thread, i posted the numbers for Milan as well...

Page 159 - Discussion - Zen 5 Discussion (EPYC Turin and Strix Point/Granite Ridge - Ryzen 8000)

Page 159 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[HurleyBird]

For IPC too, lol.

Can you corroborate this claim at all?

 

[adroc_thurston]

Can you corroborate this claim at all?

Zen3 ISSCC presentation slide where AMD talks about their IPC/Cac ratio.

 

[H433x0n]

what is your rough IPC range for Zen 5 if it is more than 22%

He’s said 32% Specint 1T in the past, which seems dubious now.

I personally anticipate it to be 18% +/- 2%. Similar to the results of Zen 3 but this time client desktop won’t see a boost in clocks. How well it performs will come down to clocks they can achieve (at least for desktop).

 

[adroc_thurston]

He’s said 32% Specint 1T in the past

Well it wasn't 1t but you only have to wait for so long!

 

[Markfw]

Well it wasn't 1t but you only have to wait for so long!

OK, so how long until we get some real benchmarks, as in NDA expired, server or desktop, I don't care.

 

[adroc_thurston]

so how long until we get some real benchmarks

Until launch.

 

[H433x0n]

Well it wasn't 1t but you only have to wait for so long!

What was it then? I may have phrased it wrong when I wrote 1T.

What I meant was IPC at iso clocks. So a Zen 4 core and Zen 5 core locked at ~4ghz with the Zen 5 core being 32% faster.

 

[Markfw]

Until launch.

OK, and that is when ????

 

[adroc_thurston]

So a Zen 4 core and Zen 5 core locked at ~4ghz with the Zen 5 core being 32% faster.

Yeah.

OK, and that is when ????

Late March or April.

 

[H433x0n]

Yeah.

Hopefully you’re still here in March / April to either take a victory lap or eat crow.

I’m hoping your information is from Zen 5 desktop samples. There’s been remarkably little leaked performance info from a desktop chip that’s launching 6 months from now.

 

[adroc_thurston]

Hopefully you’re still here in March / April to either take a victory lap or eat crow.

Well the perf is pretty reliable (if hardly believable) numbers but I'll probably eat crow wrt Granite Ridge pricing.

 

[Joe NYC]

Well the perf is pretty reliable (if hardly believable) numbers but I'll probably eat crow wrt Granite Ridge pricing.

That would be the best of both worlds.

IMO, AMD needs to use the attractiveness of Zen 5 to try to gain market share (client and server).

 

[adroc_thurston]

AMD needs to use the attractiveness of Zen 5 to try to gain market share (client and server).

Marketshare is irrelevant if you're forcing a price war against an IDM and tanking the margin.