Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

Page 831 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Josh128

Golden Member
Oct 14, 2022
1,319
1,985
106
It appears the 16 core CCX rumor was real. It appears to have the same structure of the 8 core CCX but doubled. This is why the CCDs are long and rectangular in 9965 while not so in 9754. You can see the different CCX layout clearly vs Bergamo.

...........................128C Zen 5.....................................................196C Zen 5C

Turin.png

................................................128C Zen 4C

1728596889498.png
 
Last edited:

Joe NYC

Diamond Member
Jun 26, 2021
3,647
5,177
136
Few points:

* Turin with that 8000+ RAM would surely shine even more. Too bad AMD went the cheap road again.
* 400W TDP sounds horrific but this headroom allows to get those many core gains
* AFAK AMD slides did not compare to GNR but still Turin defeats it rather easily
* Does anybody remember those slides presenting challenging routing of Rome's 8 CCDs? Today Turin got 16 of those...

DDR5-6400 is a JEDEC standard, but MrDIMM-8000 is not a JEDEC standard. It is a custom Micron implementation.

I very much doubt the OEMs / cloud providers will be very interested in adopting a non-standard memory when the standard is just around the corner. For now, it is mainly a "benchmark special"
 

inquiss

Senior member
Oct 13, 2010
511
763
136
Maybe it’ll change with Zen6, if Zen6 DT will share die with Zen6 mobile (instead of server) like some have speculated.

In that case, perhaps one should skip Zen5 on DT and wait for Zen6.
I find myself agreeing with one of your posts! I agree, zen 6 being merged with mobile could mean a bit more of a focus on consumer workloads and optimisation.

However, that might be more in the uncore and packaging etc than any significant core differences.
 

Mahboi

Golden Member
Apr 4, 2024
1,058
1,969
96
Some screens
PfSU79AozsZfgaEQ3Sx764-970-80.png.webp
fhwk5McQZaPGNTeESUwgz3-970-80.png.webp
AMD just loved Intel's SW.
Like a sailor.
 

CouncilorIrissa

Senior member
Jul 28, 2023
724
2,680
106
So was there any mention of the fabled new 16 core CCX in the 3nm 9965 chip? I dont recall any talk at all about it.
It was mentioned in the slide deck.
1728598245632.png
I hope they release Threadripper cpu based on Zen 5 sometime next year because based on these benchmarks they will be insanely fast
And just as insanely expensive.
 

jpiniero

Lifer
Oct 1, 2010
16,804
7,251
136
When you are a corp, efficiency and performance outweigh pricing so much its almost not worth considering

Keep in mind that Enterprise doesn't normally buy the top model, far from it. Intel's discounts (up to now) might be that massive to make it somewhat competitive if you didn't want to spend too much money. And Intel could only do this because 10 nm is that cheap.

Turin might be good enough to change that.
 
  • Like
Reactions: Tlh97 and Joe NYC

Josh128

Golden Member
Oct 14, 2022
1,319
1,985
106
I want to see some R23 and R24 nT benchmarks, lol. Are either of those limited to 256 threads?
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
27,241
16,107
136
I want to see some R23 and R24 nT benchmarks, lol. Are either of those limited to 256 threads?
I think that 128 threads would prove what you want to know, but from the existing benchmarks, there are no weakness in Turin.
 

H433x0n

Golden Member
Mar 15, 2023
1,224
1,606
106
🤣🤣🤣🤣 wipeout!
It does win by a decent margin but some of this is due to GNR issues. The 2P scaling for GNR is just 1.2x, whereas for Turin it’s ~1.4x-1.5X, same is true with earlier Xeon like Emerald Rapids so it’s something uniquely broken with GNR at the moment.

This set of benchmarks also included NAMD, whereas the original 6980P review didn’t include it because it’s currently bugged on GNR (unless using oneAPI), this is why the geomean for GNR was reduced by 10% in this comparison compared to the day 1 reviews. So the bugged NAMD benches drags down the geomean for 1P as well as the meager 1.2X scaling for 2P exaggerating things too.

If you compare Epyc 9755 and Xeon 6980P in 1P configuration, at the moment Turin has a lead of 18.3% in the geomean of tests (this is with the bugged NAMD & miniBUDE results).

Michael Larabel’s comments below:

Xeon 6980P does have some odd scaling with 2P / performance issues with 2P if looking at a few of the benchmarks like NAMD.... Intel was aware and reproduced my original review data and was investigating since launch but haven't heard anything more from them (granted there's staffing changes, etc, going on there). And the GNR 1P / 2P behavior did reproduce with both DDR5-6400 and MRDIMMs as you can see on the geo mean.

That's why I left e.g. NAMD out of my original GNR review to give Intel time for feedback/guidance. Like on the NAMD side they reproduced but then recommended I use the oneAPI compiler for better performance. Even though on every other CPU tested I was using the official NAMD binaries each time and behaving as expected. In the two weeks since no further updates and to provide EPYC insight into NAMD and other areas, the tests were included as that's what can be observed right now on the platforms when running the tests the same.
 

Josh128

Golden Member
Oct 14, 2022
1,319
1,985
106
It does win by a decent margin but some of this is due to GNR issues. The 2P scaling for GNR is just 1.2x, whereas for Turin it’s ~1.4x-1.5X, same is true with earlier Xeon like Emerald Rapids so it’s something uniquely broken with GNR at the moment.

This set of benchmarks also included NAMD, whereas the original 6980P review didn’t include it because it’s currently bugged on GNR (unless using oneAPI), this is why the geomean for GNR was reduced by 10% in this comparison compared to the day 1 reviews. So the bugged NAMD benches drags down the geomean for 1P as well as the meager 1.2X scaling for 2P exaggerating things too.

If you compare Epyc 9755 and Xeon 6980P in 1P configuration, at the moment Turin has a lead of 18.3% in the geomean of tests (this is with the bugged NAMD & miniBUDE results).

Michael Larabel’s comments below:
If you compare the EPYC 9965 and Xeon 6980P in 1P config, you get the same +18.3% perf in addition to 20% less average power usage. Add those two up and its a landslide win for AMD. Add to that the fact that 9965's MSRP is around 20% cheaper than the 6980P and Intel is not even in the picture anymore.
 

DavidC1

Golden Member
Dec 29, 2023
1,833
2,960
96
Add to that the fact that 9965's MSRP is around 20% cheaper than the 6980P and Intel is not even in the picture anymore.
Phoronix also has many benchmarks where it doesn't scale well with amount of cores, and in the ones that do, and more representative of a server test, even Emerald Rapids got annihilated, while GNR is now within the ballpark.

It is still a big win for AMD, but like many of us said, the closest it has been in years. The platform/software problems whether due to immaturity or other reasons are unfortunate.

It will make a big difference for Intel as before they could only target at best 50% of AMD's stack. Now they can do 70-80%.
 

H433x0n

Golden Member
Mar 15, 2023
1,224
1,606
106
If you compare the EPYC 9965 and Xeon 6980P in 1P config, you get the same +18.3% perf in addition to 20% less average power usage. Add those two up and its a landslide win for AMD. Add to that the fact that 9965's MSRP is around 20% cheaper than the 6980P and Intel is not even in the picture anymore.
IMO, It's hard to say anything definitive since it seems Intel rushed out GNR half baked to pre-empt the Turin launch. It appears that has backfired on them because it's making their platform look worse from various bugs. I can't give numbers for GNR 1P with the bugged tests removed since annoyingly Phoronix didn't include any 1P results in their day 1 review of GNR. However, If you remove the bugged NAMD tests in the 2P results, the geomean goes from 1006 -> 1106, roughly +10% boost. I don't know if 1P would get the same +10% boost or not, or if it effects the perf/watt numbers in any meaningful way either.
 

DavidC1

Golden Member
Dec 29, 2023
1,833
2,960
96
IMO, It's hard to say anything definitive since it seems Intel rushed out GNR half baked to pre-empt the Turin launch. It appears that has backfired on them because it's making their platform look worse from various bugs. I can't give numbers for GNR 1P with the bugged tests removed since annoyingly Phoronix didn't include any 1P results in their day 1 review of GNR. However, If you remove the bugged NAMD tests in the 2P results, the geomean goes from 1006 -> 1106, roughly +10% boost. I don't know if 1P would get the same +10% boost or not, or if it effects the perf/watt numbers in any meaningful way either.
I have answered you in the Intel thread.
 
  • Like
Reactions: H433x0n

gdansk

Diamond Member
Feb 8, 2011
4,568
7,681
136
Are there any efficiency numbers per core with Zen 5 on N3 vs Zen 4 server chips? I am looking for the power efficiency gains of TSMC 3nm.
If you're trying to find the impact of process alone shouldn't one compare the Zen 5C on Strix Point (N4P) to Zen 5C in Turin (N3E)? Well, I guess there's no reasonably comparable configuration and they have different core complexes.
 

Hans de Vries

Senior member
May 2, 2008
347
1,177
136
www.chip-architect.com
Last edited:

Joe NYC

Diamond Member
Jun 26, 2021
3,647
5,177
136
5.2 GHz all core for the 9800X3D. Gaming at 5.3 to 5.4 GHz would be a 6% to 8% higher frequency as the 7800X3D.

So possibly a 10% higher gaming performance as the 7800X3D.



The all core boost for the 7800X3D is ~4.8 GHz here:



10% is a lot on a bar chart.

View attachment 109189

I wonder what the official boost clock will be.

5.2 GHz while running all cores loaded with Cinebench is very good, but boost clock typically apply to 1 or 2 cores.