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

[DisEnchantment]

Speculate at will

 

[poke01]

I think the reviews will look more positive for the 9950x

 

[del42sa]

9700X PBO Performance Gained and Efficiency Hit - is PBO worth it?​

 

[Det0x]

9700X PBO Performance Gained and Efficiency Hit - is PBO worth it?​

Jay is so novice on the AM5 platform that it hurts my brain watching these videos from him, he have no idea how to setup/tweak these platforms
And instead of using Hynix A-die / M-die like everyone else, he is running some subpar d-die micron memory @ 5600MT/s at so bad timings its not even funny..
These are the timings used ? (i run tighter timings than this at 8800MT/s on a-die)

Gods knows that FCLK he is using for the reviews?

 

[PJVol]

Jay is so novice on the AM5 platform that it hurts my brain watching these videos from him, he have no idea how to setup/tweak these platforms

Wait..., curve optimizer settings for Zen5 are now in mv's, not in VCO counts?

 

[H T C]

9700X PBO Performance Gained and Efficiency Hit - is PBO worth it?​

Stupid question: has anyone seen 9700X PBO being tested WITHOUT it being "close to the max"? Or does PBO only have a "close to the max" mode?

The reason i'm asking is because i think it would obviously perform better than @ stock while NOT throwing efficiency away completely.

 

[poke01]

Gigabyte Technology Co., Ltd. X670E AORUS MASTER - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. X670E AORUS MASTER with an AMD Ryzen 9 9950X processor.

browser.geekbench.com

here it is stock 9950x in gb5

 

[Det0x]

Wait..., curve optimizer settings for Zen5 are now in mv's, not in VCO counts?

More misinformation.. Its CO values not undervolt mv, just like you say
(ASRock bios is wrong)

I recommend these two videos from skatterbench as beginner CO tuning guide

These are guides are still missing some information, such as cores close to each other are borrowing/stealing power from each other, on a per core CO curve --> example CO value on core 4 is affecting boosting on core 3.

But still, these are leagues above than anything else any youtubers have published so far 👍

I write this on a other forum, but i can post here also i guess:

Maybe this will help you understand the different bands of curve shaper better

Between the different temperature bands -5, 50 and 90 the values are extrapolated
I use it as a final trim after i have completed my normal CO curve.. Normally you have to make sacrifices in regard to your curve..

If you have tuned for max MT performance with CO, often your cpu can be unstable at max frequency / or you have to limit your fmax
Or when you have tuned your system for stability at max frequency, it can happen you are leaving headroom on the heavy MT workloads (think linpack extreme for example)
Now with the curve shaper you optimize the region you normally would have to leave untapped headroom in

Now more concrete how i would do it:
First i start working on my curve for ST workloads with apps such as GB6, corecycler and 3dmark profiler.
After iam done with that i will try to lower the CO values with shaper in lets say "medium frequency" and "high frequency", in the mid to high temp bands until system starts failing in MT workloads such as y-cruncher, linpack and OCCT large etc. After system starts locking-up/bluescreening in the MT workloads you can back alittle up and hopefully have a fully stable system.

This is feature new for all of us, so this will take some time figuring out the best way.
And as we know, the forum hivemind works best when we are all sharing information

 

[Hotrod2go]

Jay is so novice on the AM5 platform that it hurts my brain watching these videos from him, he have no idea how to setup/tweak these platforms
And instead of using Hynix A-die / M-die like everyone else, he is running some subpar d-die micron memory @ 5600MT/s at so bad timings its not even funny..
These are the timings used ? (i run tighter timings than this at 8800MT/s on a-die)
View attachment 105150
Gods knows that FCLK he is using for the reviews?

Exactly my thoughts after watching the vid. Does he do this on purpose cause' he's a closet Intel guy? who knows! but just like you say there's some easy optimizations he could have done before hand it make you think..
Not a fan of that Ryzen Master app either, call me old school but I do all my OC tests & profiles from bios only.
I have no doubt Zen 5 will get better & better with bios maturity & agesa updates as is per usual with any new cpu microarchitecture.

 

[tsamolotoff]

. Does he do this on purpose cause' he's a closet Intel guy? who knows!

Nah, he's just an average guy who doesn't really delve into things and does low-effort OC etc. Pretty much no english-speaking reviewer actually tunes memory or does proper testing that minimizes the influence of GPU (in games) or IO (in software) on the test results. On the other hand, that's what the average Joe/Jay (pun intended) would get when he tries to tinker with his newly bought Walmart PC etc (provided Walmart didn't cheap out on MB (and not just on memory) so it actually supports some tuning instead of none).

 

[Rheingold]

He also says that despite 65W TDP the CPU goes up to 88W "for bursty loads", which show a complete lack of understanding for how power limits on Ryzen work, but...

he is running some subpar d-die micron memory @ 5600MT/s at so bad timings

...this might actually be a blessing in disguise because with that lower clocked memory we get to see 4.65 GHz all-core @88W where most other publications got around 4.4 GHz. I see a "CPU Power" of 71W in Ryzen Master which by looking at some combined screenshots in Image Search seems to adhere pretty close to the "CPU Core Power" in HWiNFO and that's just around 60W for those other publications, for example 58W at ComputerBase. ComputerBase didn't test CB R23 but the 20,524 are fittingly 3-5% higher than some of the lower results in reviews.

That shows how sensitive the results are: use fast memory, IOD uses more power, lose some watts for the core power, get low scores in benchmarks that don't rely on memory speed. It's crazy that the power budget for the cores can vary by more than 20% just by using different RAM.

 

[whoshere]

I've not seen any reviews testing Zen 5 memory scaling performance but W1zzard from TechPowerUp has told me that it's coming.

From what I've heard/understood Zen 5's relatively "poor" performance uplift could be explained by its IO die that hasn't seen any improvements for this generation and was taken from Zen 4 as is.

 

[Det0x]

I've not seen any reviews testing Zen 5 memory scaling performance but W1zzard from TechPowerUp has told me that it's coming.

From what I've heard/understood Zen 5's relatively "poor" performance uplift could be explained by its IO die that hasn't seen any improvements for this generation and was taken from Zen 4 as is.

Zen5 single CCD memory bandwidth is actually worse than Zen4 at same settings, while dual CCD Zen5 is better than Zen4, as strange as it may sound..

8core single CCD Zen4 @ 6600/2200 1:1 (v-cache dont affect the memory bandwidth numbers)


8core single CCD Zen5 @ 6600/2200 1:1


I cannot show numbers for 16 core Zen5 yet

 

[Heartbreaker]

Yeah, I remember that. Yet HUB missed this fact in their last video - 17:33 , where Tim talked about Zen5 launch issues, and that's doing them no credit.

Why would HWUB be responsible for what AMD told someone else? He's going by the Information he has from AMD directly, and it's right there in your link, where you can hear the AMD's David McAfee talking over the slides saying "if you look at this compared to 7800X3d ... You'd see a couple of percentage points advantage for the 9700X."

 

[Det0x]

Nice writeup from the dev of y-cruncher here

Zen5's AVX512 Teardown + More...

Recommend reading the full write-up linked above

 

[del42sa]

Does he do this on purpose cause' he's a closet Intel guy? who knows! but just like you say there's some easy optimizations he could have done before hand it make you think..

of course he is doing it on a purpose, he´s secret Intel agent and so did HWU with their testing , I am getting feeling, that everybody want´s to hurt poor AMD

 

[In2Photos]

HWUB with a post reflecting on previous gen releases.

https://twitter.com/x/status/1822941470594363490

 

[Josh128]

Zen5 single CCD memory bandwidth is actually worse than Zen4 at same settings, while dual CCD Zen5 is better than Zen4, as strange as it may sound..

8core single CCD Zen4 @ 6600/2200 1:1 (v-cache dont affect the memory bandwidth numbers)
View attachment 105161

8core single CCD Zen5 @ 6600/2200 1:1
View attachment 105160

I cannot show numbers for 16 core Zen5 yet

Do these objectively awesome RAM numbers translate to any kind of real gaming / perf gains for Zen 5? Any crazy Geekbench gains? What kind of difference do you you see vs the standard EXPO DDR 6000 CL30 settings?

 

[CakeMonster]

Zen5 single CCD memory bandwidth is actually worse than Zen4 at same settings, while dual CCD Zen5 is better than Zen4, as strange as it may sound..

My interest is piqued. I doubt that makes a ton of difference in gaming, but I guess one advantage of the Youtube drama stuff going on is that this will be analyzed to death.

I was dead set on getting a 9950X, because even less of an improvement than Z4 over Z3 would be fine to me (and I hate the heterogenous thread prioritization schemes in the 2x CCX X3D models), but the results so far are just not good enough to even consider it. If the 2 CCX parts can claw back several percentages, maybe more people will consider getting them after all.

 

[Det0x]

Do these objectively awesome RAM numbers translate to any kind of real gaming / perf gains for Zen 5? Any crazy Geekbench gains? What kind of difference do you you see vs the standard EXPO DDR 6000 CL30 settings?

L1 tech has 2133/6400 vs 5600 in of the 1080p game-charts
But these are also just junk EXPO timings vs EXPO timings

But as ive said many times already, if your looking for a gaming cpu --> wait for X3D

 

[del42sa]

L1 tech has 2133/6400 vs 5600 in of the 1080p game-charts
But these are also just junk EXPO timings vs EXPO timings

But as ive said many times already, if your looking for a gaming cpu --> wait for X3D

no wait for ZEN6 X3D²

 

[Fjodor2001]

1). No gaming on corpo desktops. Also probably no encoding/transcoding, Photoshop, or 3d rendering. Maybe PS but that really depends on the work being done.
2). Lots of Office365
3). Lots of in-house apps which were probably cobbled together using electron

Not so obvious to me what the difference would be between office vs home PC, except for the gaming part. But not everyone that uses home PC use it for gaming either. Photoshop / 3D rendering is not so common in home either, at least not much more common than in office.

Most office type work is also what home PC is used for. Web browsing, MS Office, etc.

So I think there should be quite a lot of overlap of use cases between home vs office PC. Also, the question would anyway be what this boils down to w.r.t. actual suitable CPU arch differences, for the few differences that remain.

 

[PJVol]

Why would HWUB be responsible for what AMD told someone else?

Responsible ...? He could have just taken it into account at least, and why not if that "someone" is the TH managing editor? Besides it's no more reckless than interpreting

"if you were to look at this [the chart comparing 9700X vs. 5800X3D in a number of cherry-picked games ] versus the 7800x3d uh again an AM5 processor in um in the Zen 4 generation you'd see a a couple percentage points advantage for the 9700X over the 7800x3d."

as "the 9700X is 2% faster than the 7800X3D" without any supporting evidence provided.

PS: However, this in no way detracts from the brilliance of yet another example of AMD's notoriously stupid marketing.

 

[Hail The Brain Slug]

Ryzen 7 9700X im Linux-Test: Schneller als unter Windows? [Update]

Nachdem der Ryzen 7 9700X und Ryzen 5 9600X unter Windows nicht ganz das halten konnten, was AMD versprochen hat, muss sich der 9700X unter Linux beweisen.

www.pcgameshardware.de

They only test a handful of games but every game has a performance improvement in lows and averages on Linux with Proton vs Windows 11. An issue with scheduling is beginning to look more likely to me as a major culprit for the performance differential from official figures to actual reviews.

 

[gdansk]

Ryzen 7 9700X im Linux-Test: Schneller als unter Windows? [Update]

Nachdem der Ryzen 7 9700X und Ryzen 5 9600X unter Windows nicht ganz das halten konnten, was AMD versprochen hat, muss sich der 9700X unter Linux beweisen.

www.pcgameshardware.de

They only test a handful of games but every game has a performance improvement in lows and averages on Linux with Proton vs Windows 11. An issue with scheduling is beginning to look more likely to me as a major culprit for the performance differential from official figures to actual reviews.

It's weird. I was looking through lkml for zen related changes for months. I find no changes in lkml that mention scheduling to accommodate Zen 5 or znver5 or any code names I knew of.

So if it is a problem with scheduling, it seems whatever Linux was already doing was just fine for Zen 5.

Edit: And they should compare to 7700 in Linux too.

 

[inquiss]

Ryzen 7 9700X im Linux-Test: Schneller als unter Windows? [Update]

Nachdem der Ryzen 7 9700X und Ryzen 5 9600X unter Windows nicht ganz das halten konnten, was AMD versprochen hat, muss sich der 9700X unter Linux beweisen.

www.pcgameshardware.de

They only test a handful of games but every game has a performance improvement in lows and averages on Linux with Proton vs Windows 11. An issue with scheduling is beginning to look more likely to me as a major culprit for the performance differential from official figures to actual reviews.

I've been thinking the same thing. Microsoft seems to have focused a bunch of their resources in WoA instead of windows for x86. Pure speculation but Linux tests so much better, I can't think of another reason. Hope CnC does a deep dive on gaming performance. It's a real conundrum what's holding zen5 back.