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

[DisEnchantment]

Speculate at will

 

[IEC]

Level1 has also launched its 9950X Linux gaming (p)review:

 

[gdansk]

ouch, you guys seen that the 7950X performs better in the PS3 emulation than the 9950X? In the TPU review.

This was predictable when Yee showed that there is a regression in 128 bit op latency and RPCS3 uses pretty much only 128 bit encodings of AVX512 ops.

 

[branch_suggestion]

AMD’s Ryzen 9950X: Zen 5 on Desktop

AMD’s desktop Zen 5 products, codenamed Granite Ridge, are the latest in the company’s line of high performance consumer offerings. Here, we’ll be looking at AMD’s Ryzen 9 9…

chipsandcheese.com

 

[Glo.]

So the copium and hopium that we currently have is that its three things:

Not enough memory bandwidth for those CPUs, not redesigned IOD results in higher latency, and Windows being Windows, as usual?

 

[gdansk]

So the copium and hopium that we currently have is that its three things:

Not enough memory bandwidth for those CPUs, not redesigned IOD results in higher latency, and Windows being Windows, as usual?

Well, I don't think it is copium to say Windows sucks. But new CPUs should be able to run even garbage software faster. That's the whole point of new CPUs.

 

[Josh128]

Let me help those still suffering from copium/hopium. Let it go. Todays reviews have completely killed my desire to upgrade from my Zen 3 riggy. I have made peace with that.

DUMPSTER
FIRE

 

[yottabit]

Let me help those still suffering from copium/hopium. Let it go. Todays reviews have completely killed my desire to upgrade from my Zen 3 riggy. I have made peace with that.

DUMPSTER
FIRE

View attachment 105288

Right? I may consider a 9600 as an upgrade for a ZFS file / Plex server but… I don’t think I’ll be replacing my 5950x with anything Zen5. Although it will be interesting how things improve over the year with AGESA and Windows Scheduler improvements

I hold out some small hope for the 9800X3D specifically as a gaming CPU but we’ll see.

 

[Glo.]

At this point, I want a MiniPC like Mac Studio with Strix Halo.

 

[Rheingold]

Core-to-Core latencies across CCD boundaries more than doubled!?



WTF is going on here?

 

[Rheingold]

stupid core parking feature

JayTwoCents of all the outlets did an analysis of this and provides a fix.

 

[StefanR5R]

When 16% is the average, there should be just as many results over 16% as there is under 16%.

Nitpick: Geometric Mean and Average and Median are not the same.

"When 16% is the median, there should be just as many results over 16% as there is under 16%" would be correct.
AMD however computed the geometric mean over their selection of tests.

(Besides, they wrote "*all results are 'up to'. See endnote GNR-03." Furthermore, always keep absolute performance and iso-clock performance and clock-normalized performance and IPC apart; they are all different things. AMD's slide showed relative iso-clock performance.)
________
edit: @yottabit was way ahead of me. :-)

 

[tsamolotoff]

Core-to-Core latencies more than doubled!?

This is close to the penalty of accessing the other CPU ram in an 2S system, very weird, since io die is the same it shouldn't have changed.

 

[Tuna-Fish]

Core-to-Core latencies across CCD boundaries more than doubled!?

View attachment 105289

WTF is going on here?

This is horrifying. It easily explains the regressions.

 

[igor_kavinski]

Across a very large test suite, an 18% performance improvement while using 3.5% less power is a very solid generational improvement. No category of benches saw less than like 12% improvement (gaming not tested). Linux users rejoice, I guess, you got a good upgrade, lol.

But even in Linux, it's concerning that the 9900X is beating the 9950X in some benchmarks. That should NOT happen. Either the extra cores are getting membw starved or hitting power/temp limits.

 

[sl0519]

Screenshot from Bilibili in 360p, sorry for the bad clarity.
It seems like performance uplift is there, so why is the avg application performance showing single digit uplift according to Techpowerup?

 

[igor_kavinski]

JayTwoCents of all the outlets did an analysis of this and provides a fix.

What does the cache optimizer driver have to do with a non-X3D CPU???

 

[Hitman928]

View attachment 105293

Screenshot from Bilibili in 360p, sorry for the bad clarity.
It seems like performance uplift is there, so why is the avg application performance showing single digit uplift according to Techpowerup?

It seems to be similar to the 9700X in that the power has been limited to a lower value than Zen 4. With the 9950x, the TDP is the same but the PPT is reduced from 230 W to 200 W, so the real power consumption is lower for the 9950x, reducing its performance improvement but increasing its efficiency improvement. HWUB still seems to be the odd man out in this regard, but the others show good efficiency improvements.

 

[coercitiv]

Whelp, we have confirmation the review guide was a mess. They had to message reviewers later and tell them the dual CCD parts need special attention. You can click on the image to open timestamped GN 9950X review.

 

[Hitman928]

Core-to-Core latencies across CCD boundaries more than doubled!?

View attachment 105289

WTF is going on here?

I'm guessing it's related to the new core parking driver which is turning off the cores on 1 CCD. This has the effect of reducing lightly threaded power consumption, see the improvements shown in TPU's testing. Theoretically it should also help gaming performance by keeping all the game threads on 1 CCD, but I think that has been a mixed bag according to reviewer numbers, but who knows if some reviewers have the driver installed correctly or not.

 

[Mopetar]

With all of these mixed results, I think I'm going to hold out for the userbenchmark review before making any decision.

 

[tsamolotoff]

I'm guessing it's related to the new core parking driver which is turning off the cores on 1 CCD.

Not sure about driver, but I remembered that AT tests the CPUs with JEDEC memory, so the IO die might be in some deep sleep mode that disables the GMI links to secondary CCD, at least I hope so. This is like first gen 32 core Threadripper-grade latency chasm, it'd be absolutely pointless to buy 9950x3d if this thing is not fixed

 

[igor_kavinski]

Don't know why AMD insists on crapping their hardware launches. It's like they demand it.

"We absolutely MUST have some launch drama!"

Said Queen Lisa.

And all her courtiers nodded with glee.

 

[SolidQ]

y AMD insists on crapping their hardware launches

they should call it just Zen4+, like Zen1 and Zen1+

 

[Hitman928]

Not sure about driver, but I remembered that AT tests the CPUs with JEDEC memory, so the IO die might be in some deep sleep mode that disables the GMI links to secondary CCD, at least I hope so. This is like first gen 32 core Threadripper-grade latency chasm, it'd be absolutely pointless to buy 9950x3d if this thing is not fixed

If that was the case, then Zen 4 should show the same behavior, but it’s unique to Zen 5.

 

[igor_kavinski]

they should call it just Zen4+, like Zen1 and Zen1+

No, in many cases it's Zen 4 minus minus.

Everybody, do yourself a favor and DON'T check out the igorslab.de 9950X review.

You don't deserve depression. No one does.

Don't GO THERE!