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

Page 158 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

DisEnchantment

Golden Member
Mar 3, 2017
1,586
5,694
136
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 :D)

We already got roadmaps and key technologies like AIE
1664493390795.png

1664493471118.png

1664493514999.png

1664493556491.png
1681912883215.png
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
 
Last edited:

H433x0n

Senior member
Mar 15, 2023
822
870
96
For such a radical design, I would expect a lot more than 15% gain to be worth the risks and power usage and not to mention possible security implications.

That’s my problem too, I can’t reconcile it in my head that such a massive change only amounts to an IPC increase of 10-15%. Especially if there’s a rumored clock regression for client desktop. Although for Turin I don’t think there’s a clock regression since I’ve seen rumors of increased clocks. Turin would be fine even in worse case scenario where it gets 15% IPC increase if they bump clocks by 1 or 2 speed bins.

Everybody points to AMD under estimating IPC for Zen 4 but that was in marketing material where I assume they wanted to play it safe. I really wish we could see the internal estimates for Zen 3 & 4 pre-launch to get an idea of how this went in the past.
 
Last edited:

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,394
14,356
136
@adroc_thurston What again is your best guess for Zen 5 desktop availability (to buy) ? I have a motherboard that I was going to use for a dual 4090 build, but I may hold off it this will be around say Jan or Feb 2024.
 

adroc_thurston

Golden Member
Jul 2, 2023
1,661
1,875
96
I can’t reconcile it in my head that such a massive change only amounts to an IPC increase of 10-15%
a) that's nT IPC (basically per-vCPU iso clk perf)
b) you're getting more!
I really wish we could see the internal estimates for Zen 3 & 4 pre-launch to get an idea of how this went in the past.
7-10% (got 13%) and 8% (got 14%).
What again is your best guess for Zen 5 desktop availability (to buy) ?
Late March or April.
 

Abwx

Lifer
Apr 2, 2011
10,822
3,286
136
Think he is talking SPEC int Rate for server parts, which my quick google gives about 13% for zen 2 to 3 , i cant find good zen 3 to 4 numbers.
assuming the theme keeps from the original MLID image

Numbers from AT, although it was less than 19% in Cinebench...

IPC wise, looking at a histogram of all SPEC workloads, we’re seeing a median of 18.86%, which is very near AMD’s proclaimed 19% figure, and an average of 21.38% - although if we discount libquantum that average does go down to 19.12%. AMD’s marketing numbers are thus pretty much validated as they’ve exactly hit their proclaimed figure with the new Zen3 microarchitecture.

 

Abwx

Lifer
Apr 2, 2011
10,822
3,286
136

That s surely some average like these ones :

26-1280.0e92feb1.png



3-1280.4e3998b4.png

 

Abwx

Lifer
Apr 2, 2011
10,822
3,286
136
That's not server.
Christ.

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.

It was indeed 14% for servers and 13% for DT when it comes to Zen 4, but this is close enough that they could retain the higher number for the purpose of exhibiting the higher possible one on the slide.

 

adroc_thurston

Golden Member
Jul 2, 2023
1,661
1,875
96
but this is close enough that they could retain the higher number for the purpose of exhibiting the higher possible one on the slide.
No, it's just whatever they have for 1st party measurements in server, and they never presented nT IPC for Milan since it sucked so they plugged the client number in.
 

Abwx

Lifer
Apr 2, 2011
10,822
3,286
136
No, it's just whatever they have for 1st party measurements in server, and they never presented nT IPC for Milan since it sucked so they plugged the client number in.

For Zen 3 they were stuck with DDR4 wich didnt help for high core count SKUs but was good enough for DT CPU, this time numbers will be more favourable for DT CPUs as well.

That being said i dont believe much in 10-15% IPC improvement for Zen 5, if the uarch slide is accurate then it should be something like 20-22%.

Historically they improved IPC by 52% with 100% more ALUs, with 50% more the number is unlikely to be as low as the 10-15% that was leaked, Intel got as much by increasing the ALU count by a mere 33%, 3 ALUs to 4, from SB/IBridge to Haswell.
 

adroc_thurston

Golden Member
Jul 2, 2023
1,661
1,875
96
For Zen 3 they were stuck with DDR4 wich didnt help for high core count SKUs but was good enoygh for DT CPU
Not the issue, the thing just had negative PPW over Zen2 in silly nT loads.
if the uarch slide is accurate then it should be something like 20-22%.
Those were my expectations, turns out to be more.
Historically they improved IPC by 52% with 100% more ALUs, with 50% more the number is unlikely to be as low as the 10-15% that was leaked, Intel got as much by increasing the ALU count by a mere 33%, 3 ALUs to 4, from SB/IBridge to Haswell.
ALU count is a rather silly metric.
X4 has 8 of them and isn't that much faster over 4ALU X1.
 

Abwx

Lifer
Apr 2, 2011
10,822
3,286
136
Not the issue, the thing just had negative PPW over Zen2 in silly nT loads.

5950x clocked 3.3% higher in MT than the 3950X, just this clock uplift will eat 5% perf/Watt comparatively.
Overall the former still consumed 9W less for 15% better perfs in, Mt ,so no, perf/watt was improved overall contrary to your random speculations.

All numbers can be found here :

X4 has 8 of them and isn't that much faster over 4ALU X1.

What are you talking about here..?..

There s no AMD CPU that had 8 ALUs, it was 3 with A64/Phenom and 2 for Bulldozer and then 4 for Zen.
 

JustViewing

Member
Aug 17, 2022
134
232
76
It is easier to extract more performance from FP code than x64 integer code. So 10-15% IPC for integer code seems reasonable, for FP code there is no immediate limit. Most FP code are streaming in nature, so there is high possibility to extract more IPC from FPU. The next big IPC improvement for integer code will come with new CPU design which contains support for 32 Genaral Purpose registers. With increased register count, there is more opertunity to execute parallel instructions.
 

adroc_thurston

Golden Member
Jul 2, 2023
1,661
1,875
96

Abwx

Lifer
Apr 2, 2011
10,822
3,286
136
It is easier to extract more performance from FP code than x64 integer code. So 10-15% IPC for integer code seems reasonable, for FP code there is no immediate limit. Most FP code are streaming in nature, so there is high possibility to extract more IPC from FPU. The next big IPC improvement for integer code will come with new CPU design which contains support for 32 Genaral Purpose registers. With increased register count, there is more opertunity to execute parallel instructions.
Historically INT was more difficult to improve, but let s look at Zen 4 over Zen 3, first with FP and then with INT.

CB R20 13%, CB R15 12%, Povray 12%, Blender 11%, Corona 6%.

Handbrake 15%, 7ZIP 12%, Digicortex 9%, Agisoft 5%.

Numbers for FP and INT are comparable as are the one between Zen 2 and Zen 3 in the same graph :



Milan.
7713 vs 7742 is ehhh.
I provided numbers for the 3950X/5950X, those should be comparable for server parts, overall MT perf/watt was improved by roughly 15% despite slightly higher frequency and a comparable 7nm process.

ARM Cortex-X1 and Cortex-X4.
Jesus.

Perhaps that we should stick to X86 and AMD CPUs, ARM is not a comparable design, they could prefer using more ALUs that are obviously less versatile than what is implemented in Zen.