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

[DisEnchantment]

Speculate at will

 

[adroc_thurston]

AMD basically invited Intel to eat their lunch in parallel workloads with heavy inter-thread traffic.

Like databases?

 

[StefanR5R]

IgorsLab noted "hectic" CPU clocks during testing:

"...significantly more hectic clock behavior under rapidly changing loads", said Igor Wallossek; emphasis mine. Very quick adaption of the clock to different loads should generally be regarded as a good sign.

Have we found out exactly why AMD delayed the launch? Related to the Windows/Linux disparity?

I don't now if it was a cause or just a symptom:
One thing which some reviewers noted was that there were lots of BIOS updates in quick succession close before the new review embargo date, starting off with a total inability to boot on some motherboards, and taking several iterations before arriving at a semi-reasonable state of stability. While this is what usually happens at (AMD) CPU launches, it was even worse in those reviewers view than at previous ones.

Correct me if I'm wrong, but didn't we have screenies of misprinted IHSes on Zen5 CPUs?

The photo was from an ES/QS(?), not from a retail CPU.

BTW, there are changes in the prefetch logic across L2 and L3 in Zen 5. Not sure if the motivation is to enable it to work better in different configurations but it might have added some regressions
Could be to handle bigger CCXs sharing L3 and snooping other CCXs

Could this cause these cache line bouncing tests to regress?

 

[igor_kavinski]

Like databases?

Yeah. Few examples:

Spoiler: Just horrible

The last one is a blatant example of how the inter-thread traffic is ruining the 9950X's performance, with the single CCD 9600X trampling all over it!

 

[igor_kavinski]

More:

Spoiler: Morribleu!

If AMD is unable to fix these issues, either via AGESA or a new stepping, I will have no choice but to call 9950X "broken silicon" because these regressions make no sense for the latest generation of AMD's flagship CPU.

 

[tsamolotoff]

If AMD is unable to fix these issues, either via AGESA or a new stepping,

Or this might have been a deliberate choice, as Intel at some point basically nuked legacy 'mixed-mode' avx/sse with some changes to how VZEROUPPER works. AMD clearly doesn't care about anything non-enterprise related, so they might have tweaked interconnect in a way that makes it work really bad with 'legacy' desktop IO die (and maybe it works okay with the new epyc IO die)

 

[Timmah!]

Surely this significantly higher inter-CCD latency is some software related issue and will be fixed?

Perhaps the increased latency is caused by data being run through NSA servers as they travel from one CCD to the other :-D

 

[igor_kavinski]

so they might have tweaked interconnect in a way that makes it work really bad with 'legacy' desktop IO die (and maybe it works okay with the new epyc IO die)

If these issues are also present in 9950X3D, I will be skipping Zen 5 and waiting for Zen 6.

 

[tsamolotoff]

Surely this significantly higher inter-CCD latency is some software related issue and will be fixed?

It'd be nice if some owner could have run this latency benchmark (either the one Dom used or the one CopeframeX created) in some non-Windows OS or in Win10 instead of Win11 (which all of reviewers seem to use)

PS Here's latency chart for 7950x3d, so the inter-chiplet access time increased almost 3 times which is bewildering to put it midly considering that the IO die is the same and IFOPs probably didn't change as well, at least on the physical level.

 

[igor_kavinski]

It'd be nice if some owner could have run this latency benchmark (either the one Dom used or the one CopeframeX created) in some non-Windows OS or in Win10 instead of Win11 (which all of reviewers seem to use)

Yeah, @Det0x could certainly try Win10 to see if it's a Win11 specific issue.

If confirmed, I can aim my vitriol at Microsoft instead

 

[Det0x]

Yeah, @Det0x could certainly try Win10 to see if it's a Win11 specific issue.

If confirmed, I can aim my vitriol at Microsoft instead

My earlier run was on a old win10 21H2 v19044

 

[igor_kavinski]

My earlier run was on a old win10 21H2 v19044
View attachment 105379

Is it possible to run benchmarks without installing the chipset drivers?

 

[igor_kavinski]

13% uplift in 99th percentile fps and +21% in average fps.

Looks like flightsim fans will be upgrading to the 9950X3D immediately, barring any serious issues.

 

[tsamolotoff]

Also, 5900x for comparison (fixed 4.5 ghz)

 

[sl0519]

Haven't watched it yet. Zen 5 30% IPC incoming!

 

[CouncilorIrissa]

Haven't watched it yet. Zen 5 30% IPC incoming!

Zen 3+1%

 

[Hans Gruber]

Seriously, I am not saying anything for at least a month. That will give AMD time to figure out the Zen 5 mess. I can only assume there is a AGESA bios update coming to remedy the situation.

 

[tsamolotoff]

Haven't watched it yet. Zen 5 30% IPC incoming!

Didn't watch it also, but maybe admin mode just force-disables VBS and/or HVCI ? Not sure if reviewers disable both while testing, both hurt Zen3 and onwards gaming performance substantially (like 10% in some cases)

 

[Gideon]

Haven't watched it yet. Zen 5 30% IPC incoming!

I watched it, the TL;DW is:

The workaround improves gaming performance of Zen 5 by 4%, but it also improves gaming perf by 3% for Ryzen 7xxx series, so it's more like a Ryzen specific problem, not an issue related to Zen 5 alone.

It will also be patched in windows apparently.

So doesn't really change anything regarding when comparing Zen 4 and Zen 5

 

[SK10H]

Also, tried Linpack Xtreme but it reports hardware failure within a minute or starting, think this program has not be refined for Zen 5 at this point in time.

It's a good thing you find a program that can crash your new hardware, that's the point of stress testing. Unless the program is made with illegal sequences of instruction which we know this isn't, then it's your hardware

Now go find corecycler and do AVX2 ycruncher / p95 test. Don't blame the software when your retail sample isn't QA properly.

 

[sl0519]

I watched it, the TL;DW is:

The workaround improves gaming performance of Zen 5 by 4%, but it also improves gaming perf by 3% for Ryzen 7xxx series, so it's more like a Ryzen specific problem, not an issue related to Zen 5 alone.

It will also be patched in windows apparently.

So doesn't really change anything regarding when comparing Zen 4 and Zen 5

It benefitted both Zen 4 to Zen 5. If compared to original Zen 4 data it's now a additional 5% boost to gaming performance. But since the bug is hurting both parts, the gap will not be widened even if it's patched in the future.

 

[poke01]

The most utter bull crap I heard recently from MLID.

https://twitter.com/x/status/1823938599173144758

 

[RnR_au]

The workaround improves gaming performance of Zen 5 by 4%, but it also improves gaming perf by 3% for Ryzen 7xxx series, so it's more like a Ryzen specific problem, not an issue related to Zen 5 alone.

The workaround may also work for Intel so it may not be a Ryzen issue. Noone has tested this yet.

Edit: how much tinfoil do we need if Intel is found to be unaffected by this bug.... ?

 

[branch_suggestion]

https://twitter.com/x/status/1823912091503485346

X3D launching around October also makes sense.
I personally always wait at least 3 months before buying a new product generation.

 

[yottabit]

The high latencies almost make me wonder if they were planning on doing a 16 core CCD for desktop (on N3 ?), so were planning on single-CCD only SKUs, and then had to abandon it partway through development cycle and used Elmer’s Glue to fasten two 8 core CCD together

 

[moinmoin]

so it's more like a Ryzen specific problem, not an issue related to Zen 5 alone.

It may not even be Ryzen specific, HUB just didn't test any Intel chips, only Ryzen 7700X and 9700X.

Edit: Beaten. Not sure what to add, except I found it funny to hear that Windows nowadays has an actual hidden full admin account but still gives every user plenty of admin rights.