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

[DisEnchantment]

Speculate at will

 

[tsamolotoff]

I don't know if it has been already reported.

https://twitter.com/x/status/1824038670552473709

No idea what they've did, but I've (re)enabled Admin account, run CP2077 integrated benchmark and got the same result I have on my normal (admin) account (720p with extreme upscaling enabled). But I've got VBS etc disabled as well as UAC disabled. Maybe switching to admin removed some bloatware from the auto startup or something.

 

[adroc_thurston]

Only 12% higher for 33% more threads?

PTS is a combination of 1T and nT loads.
Geomean score matters little, you need to look for specific tests.

 

[igor_kavinski]

What benchmarks do you see that the extra latency is hurting?

Let's ignore the benchmarks for a moment. I want AMD accountable for Zen 5 dual CCD CPUs showing higher inter-core communication latency. That should be reason enough to get worried that something went very, very wrong. Unless I see the "fixed" latency chart, either fixed through AGESA, Windows driver/software or new stepping, I will keep my nose up.

 

[Hitman928]

Let's ignore the benchmarks for a moment. I want AMD accountable for Zen 5 dual CCD CPUs showing higher inter-core communication latency. That should be reason enough to get worried that something went very, very wrong. Unless I see the "fixed" latency chart, either fixed through AGESA, Windows driver/software or new stepping, I will keep my nose up.

You are getting fixated on something that seems to be a non-issue. If AMD can save power while not effecting performance outside of a completely synthetic scenario that never appears in real loads, what is the issue? If additional testing shows that it can degrade performance in situations people actually care about, maybe they need to do something.

 

[igor_kavinski]

If AMD can save power while not effecting performance outside of a completely synthetic scenario that never appears in real loads, what is the issue?

AMD needs to clarify that this is what's happening and that's why synthetic benchmarks are seeing the increased latency.

Right now, I'm assuming that something in Zen 5 dual CCD CPUs is broken and I don't feel like recommending them to anyone.

 

[Hitman928]

AMD needs to clarify that this is what's happening and that's why synthetic benchmarks are seeing the increased latency.

Right now, I'm assuming that something in Zen 5 dual CCD CPUs is broken and I don't feel like recommending them to anyone.

I mean, there's not much reason to recommend Zen 5 for most users right now when Zen 4 is so relatively cheap, but that applies to both dual and single CCD products and has nothing to do with the cross CCD latency.

 

[samboy]

AMD needs to clarify that this is what's happening and that's why synthetic benchmarks are seeing the increased latency.

Right now, I'm assuming that something in Zen 5 dual CCD CPUs is broken and I don't feel like recommending them to anyone.

Agree - The need to have a special driver to park cores (i.e. disable one chiplet), that was not needed for the 7950x, is a huge red flag for me as well.

My planned 9950x build is on hold; I skipped the 7950x generation with the hope that the 9950x would give another 15-25% performance improvement. Seems to be more single digit gains for many tests and looks like I'll need to wait for the next generation. If Arrow lake gives 15-25% gains then I may need to jump ship (as much as I'm not a fan of the way Intel has handled Raptor instability).

 

[Hitman928]

Agree - The need to have a special driver to park cores (i.e. disable one chiplet), that was not needed for the 7950x, is a huge red flag for me as well.

From what we have so far, the driver does not do anything for parking the cores, that behavior is controlled by the CPU/firmware itself. The driver is there to prevent thread migration from the active CCD to the inactive CCD unnecessarily.

The 7950x wouldn't park cores on the 2nd CCD.

 

[Nothingness]

Ever consider that there may be an actual explanation for it, like it may be running into memory bottlenecks or the previously discussed anomalous results are bringing down the average?

Yes. And many apps just don't scale beyond a certain point (or even not at all), no matter whether there's a memory bottleneck or not. Lock contention might be a software issue. Even parallel compilation can be "broken" if the makefile is poorly implemented or if you're being hit by Amdahl's law where the linker does a lot of work or some of the files take a disproportionate time to compile (heavily templatized source for instance).

 

[inf64]

What I would like is AMD to come out and explain how they got the application Performance numbers they showed in the charts (which is not reproducible by anyone else right now). There is no way they achieved those numbers and everyone else is doing something wrong.

 

[Doug S]

So Linux results are generally better than Windows... yet again. Looks like even though facing head on Microsoft's abyssmal Windows scheduler since Zen 1 (remember the first Threadripper gen?) AMD still manages to surprise itself, reviewers and all of the audience how bad the scheduler really is and how completely unprepared for it AMD continues to be.

It's like the tired running gag of Lucy not letting Charlie Brown kick the ball.

I'd be curious to compare the same system when running Windows Server. Don't suppose anyone with a Zen 5 has the ability to install Server on that system as a test to see? I would assume you can install it even without a license and it'll run long enough for a few benchmarks...

It may have nothing to do with the scheduler, per se. There are things the OS controls/influences the CPU with that could be to blame. Power management is one place I'd look. Maybe Linux slants more towards the performance side of power management (more aggressive with clock ramping, which would help Geekbench) and Windows slants more towards saving power. That would make sense, given that probably 0.1% of Linux installs are on laptops, whereas over 50% of Windows (not Windows Server) installs are.

 

[del42sa]

when mobos with new AMD X870 and X870E motherboard chipsets will be released ?

 

[Ranulf]

when mobos with new AMD X870 and X870E motherboard chipsets will be released ?

Sept. 30th if I remember right. That is the rumor anyway from around Computex.

 

[Det0x]

New Apple AI benchmark released

Download Geekbench AI

www.geekbench.com

No idea if this is a good or bad score for Zen5 16core

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an AMD Eng Sample: 100-000001277-60_Y processor.

browser.geekbench.com

Cant find a single 13900k / 14900k result, crashing all these systems ? 🤣

 

[Hail The Brain Slug]

New Apple AI benchmark released

Download Geekbench AI

www.geekbench.com

No idea if this is a good or bad score for Zen5 16core

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an AMD Eng Sample: 100-000001277-60_Y processor.

browser.geekbench.com

View attachment 105414

Stock 7950X 6400/2133: https://browser.geekbench.com/ai/v1/4729

 

[igor_kavinski]

ASUS System Product Name vs Gigabyte Technology Co., Ltd. X670E AORUS XTREME - Geekbench

Oh look, there's the +35%!

 

[poke01]

The last one is the 13900k from Max.

 

[In2Photos]

AMD shouldn't go around trying to please everyone. There wasn't anything very wrong with Zen 4...

I remember the Zen 4 release very differently then. The major consensus was that Zen 4 had great performance, but hitting 95C had people screaming at the top of their lungs. It was a major focal point surrounding the release for weeks.

Part of me thinks they tried to right the ship with this release. It would have been fine if there was a performance uplift in addition to the decreased temps and power consumption. But there isn't.

 

[MS_AT]

No idea what they've did, but I've (re)enabled Admin account, run CP2077 integrated benchmark and got the same result I have on my normal (admin) account (720p with extreme upscaling enabled). But I've got VBS etc disabled as well as UAC disabled. Maybe switching to admin removed some bloatware from the auto startup or something.

Are you using Win11 or Win10? Might be it's Win11 specific issue.

Agree - The need to have a special driver to park cores (i.e. disable one chiplet), that was not needed for the 7950x, is a huge red flag for me as well.

The thing is, you don't know if there is a need as nobody has tested without the driver enabled afaik. I think they are unable to do anything about CCD to CCD latency and therefore were afraid that for low threaded games it might hurt performance. Problem is games are not created equal, some might have already had mitigations for 2 CCD latency issue [it's still 3 times bigger than inter CCD latency on Zen4] or are spawning worker threads that are doing independent tasks where limiting the number of threads will make it worse. That is why I hope reviewers will follow-up with Core Parking on/off scenario so we could see if there really is a need.

 

[yuri69]

What I would like is AMD to come out and explain how they got the application Performance numbers they showed in the charts (which is not reproducible by anyone else right now). There is no way they achieved those numbers and everyone else is doing something wrong.

The charts got the classic disclaimer attached. So the marketing ppl might have simply lied.

 

[igor_kavinski]

I remember the Zen 4 release very differently then. The major consensus was that Zen 4 had great performance, but hitting 95C had people screaming at the top of their lungs. It was a major focal point surrounding the release for weeks.

Part of me thinks they tried to right the ship with this release. It would have been fine if there was a performance uplift in addition to the decreased temps and power consumption. But there isn't.

Yes there was that. But it subsided when people actually used it and saw that their CPU didn't die and instead worked happily ever after.

 

[Hitman928]

New Apple AI benchmark released

Download Geekbench AI

www.geekbench.com

No idea if this is a good or bad score for Zen5 16core

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an AMD Eng Sample: 100-000001277-60_Y processor.

browser.geekbench.com

View attachment 105414
Cant find a single 13900k / 14900k result, crashing all these systems ? 🤣

Judging by the name, it seems like there is no AVX-512 support which is a shame. I don't know what this bench is doing, but there are multiple AVX-512 instructions specifically for speeding up AI tasks.

 

[igor_kavinski]

Judging by the name, it seems like there is no AVX-512 support which is a shame. I don't know what this bench is doing, but there are multiple AVX-512 instructions specifically for speeding up AI tasks.

Didn't see the comparison between 9950X and 7950X? That uplift isn't from a generational IPC bump.

AVX-512 is very much being used.

 

[igor_kavinski]

https://www.msi.com/blog/how-to-boost-amd-ryzen-9000-series-gaming-performance-amd-opp-msi-memory-try-It-and-high-efficiency-mode

@Det0x , is this just plain crap or some truth to it?

 

[Hail The Brain Slug]

https://www.msi.com/blog/how-to-boost-amd-ryzen-9000-series-gaming-performance-amd-opp-msi-memory-try-It-and-high-efficiency-mode

@Det0x , is this just plain crap or some truth to it?

I had an X670E Tomahawk for a brief time and it had this "High Efficiency" feature with my 7950X. I experimented with all the different settings (relaxed, balanced, tighter, tightest) and it was garbage. Inconsistent and unstable results.

Not sure about the AMD Optimized Platform Profile, that's something I haven't seen before.