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

[DisEnchantment]

Speculate at will

 

[marees]

https://twitter.com/x/status/1826555819628134553

For those without a twitter account, David continues;

View attachment 105952

Some more details

https://twitter.com/x/status/1826618084075143186

 

[fastandfurious6]

+40% IPC is achievable after all!!!!

PBO x20
disable SMT
disable windows security mitigations
windows kernel patches
fix parking behaviour
microcode updates
dual decoder fix
🙏 pray to Lisa Su morning, noon, afternoon, evening and in your sleep 🙏

 

[yuri69]

dual decoder fix

Nope, the official Zen 5 SOG confirmed the decoder works as intended - it is there to boost SMT.

 

[PJVol]

+40% IPC is achievable after all!!!!

And RDNA3 +50% ppw

 

[DrMrLordX]

Invest more on QA
Develop a better communication strategy
Ask marketing to minimize the FUD / lies

Or just keep fudging your DiY marketing while spending the real time and money on the hyperscalars, cloud companies, and major corporate buyers. AMD knows where there is money. It doesn't come from DiY.

 

[Cllaymenn]

 

[marees]

5. Problems under the hood and in the WIn 11 Kernel.. This system is still unfinished

My understanding is that Microsoft promised AMD a set of changes/optimizations for zen 5. AMD tested for these optimizations which were enabled in admin mode

Then MS did a recall & AMD fell flat on their face

But by the time 3d vcache processors are released MS should have rolled out these changes widely

 

[Cllaymenn]

My understanding is that Microsoft promised AMD a set of changes/optimizations for zen 5. AMD tested for these optimizations which were enabled in admin mode

Then MS did a recall & AMD fell flat on their face

But by the time 3d vcache processors are released MS should have rolled out these changes widely

or Intel paid MS a lot of money to introduce some "improvements" for Intel processors in the next patches before the premiere of ZEN5, which could also affect ZEN5's performance I know it's unlikely.

But maybe someone with ZEN 5 could check it by installing e.g. a Win11 build from half a year ago, disabling updates and comparing the results on ZEN 5 with the latest Win 11 release + all updates.

 

[MS_AT]

1. 2x greater delays between CCD and 7950X. (it's a bug in the core hardware or a software bug, microcode or AGESA - hopefully)

And no one yet pointed a single workload where this latency was confirmed to have noticeable influence on performance. We were shown few outliers from automatic tests, where 9950x was doing worse than 9700x but at the same time 9900x was topping the chart.

2. And I saw information that only one of these two 4-way decoders is on, and the other is off. Because something was wrong when both were working. This would explain the relatively low IPC increase in ZEN 5. Because the core was significantly expanded.

We don't know if it is by design of if there is a mistake. But profiling data from C&C show that anyway most of code is run from uop cache. If you are concerned about particular workload you would need to profile it to see if the decoder is really insufficient there. In general operation the dual decoder should be helping with general chip utilization.

3. PBO2 algorithms supposedly remained unchanged, and after all ZEN 5 differs greatly in structure from ZEN 4, so if old PBO2 algorithms try to control ZEN5 like ZEN4, it's no wonder it doesn't work well.

Any data to support the claim that there is anything wrong with boosting on Zen5? As it seems to be doing pretty well with clock management.

Young, unfinished AGESA. From generation to generation of ZEN, developing AGESA becomes more and more difficult. I don't know how many people are currently working at AMD on the AGESA code and whether this team is being strengthened, and whether the designers of the ZEN 5 arch are helping, working on this code. But they are going slowly.

Any particular thing?

Problems under the hood and in the WIn 11 Kernel.. This system is still unfinished, it has problems with supporting an efficient CPU. Phoronix tests show how big the differences are in the same calculations in WIn 11 vs. Linux. Windows 11 suppresses, wastes the efficiency of modern CPUs. Some decent AI should check the MS programmers' code, detect errors, propose optimizations of critical places in the code.

Problem is with security mitigations for older Zen CPUs that were applied to newer ones as well if we are to believe twitter reports. Now it's between M$ and AMD to solve. Could have been solved earlier. And if you think AI would be able to sensibly analyze kernel code I see that the marketing surrounding AI is really doing great

 

[MadRat]

AI like people, are only as smart as the information that trains them. And that's assuming they were intelligently designed to arrive at the correct conclusions when gleeming data. Sometimes their behaviors become corrupt by design. It will be years before we know if these AI accelerators are worthwhile.

 

[fastandfurious6]

I see it coming, when they iron out every bug and limitation

ZEN 5.5 - UNLEASH THE BEAST

promo art: Lisa Su holding the leash of Cerberus with ZEN Armor and 5.5 heads




(dunno why it made 4 heads, maybe AI knows it's the Final True Form of ZEN 4? ran out of AI credits lol)

 

[GTracing]

And RDNA3 +50% ppw

That claim wasn't nearly as unbelievable as the Zen5 40% ipc claim was.

RDNA1 increased ppw by around 25% over Radeon VII. RDNA2 by around 25-50% over RDNA1 (depending on sku). And unlike RDNA3, those two generations didn't have the benefit of a new node.

 

[Hail The Brain Slug]

That claim wasn't nearly as unbelievable as the Zen5 40% ipc claim was.

RDNA1 increased ppw by around 25% over Radeon VII. RDNA2 by around 25-50% over RDNA1 (depending on sku). And unlike RDNA3, those two generations didn't have the benefit of a new node.

Yeah but RDNA 3 +50% PPW was a direct, first party claim by AMD.

Where did AMD say 40% IPC gain for Zen 5?

 

[StefanR5R]

carefully selected test runs that produce meaningful data

In the bigger picture, it's entertainers talking about entertainment products.

 

[GTracing]

Yeah but RDNA 3 +50% PPW was a direct, first party claim by AMD.

Where did AMD say 40% IPC gain for Zen 5?

They didn't. I'm not sure I see your point. I'm saying that the 40% ipc claim (by leakers) is not comparable to the 50% ppw claim (by AMD). One was never going to happen, the other was plausible.

 

[tsamolotoff]

Cerberus

But that's Zen3

 

[Rigg]

In the bigger picture, it's entertainers talking about entertainment products.

The primary goal of a PC hardware reviewer should be to provide meaningful information to it's viewers/readers so they can make informed purchasing decisions. While yes, it's a form of entertainment for enthusiasts, it's also serves to advise people (normie and enthusiast alike) on how to best spend their hard earned money on expensive entertainment products. There is a lot of questionable PC game testing out there. It can be highly misleading especially to people who don't understand how to contextualize the data.

 

[StefanR5R]

For meaningful information about CPU performance influence on video games, they could make one video every three years.
(Edit: Better, write an article every three years.)

 

[igor_kavinski]

For meaningful information about CPU performance influence on video games, they could make one video every three years.
(Edit: Better, write an article every three years.)

You want them and us to starve (them literally and us mentally) ???

 

[MarkPost]

 

[inf64]

Outside of gaming, AMD has the best overall CPU with Zen 5, even on Windows (and especially on Linux). This doom and gloom stuff we see on the web is ridiculous. AMD made a lot of blunders, but they currently have the best products for productivity (9950X) and gaming (7800X3D and 7950X3D). Sure, it would be awesome if non X3D Zen 5 would have toped the gaming charts, but that didn't happen. Gamers will either buy 7800X3D now or wait 4 months and get Zen 5 X3D.

 

[biostud]

Outside of gaming, AMD has the best overall CPU with Zen 5, even on Windows (and especially on Linux). This doom and gloom stuff we see on the web is ridiculous. AMD made a lot of blunders, but they currently have the best products for productivity (9950X) and gaming (7800X3D and 7950X3D). Sure, it would be awesome if non X3D Zen 5 would have toped the gaming charts, but that didn't happen. Gamers will either buy 7800X3D now or wait 4 months and get Zen 5 X3D.

Zen 5 is not a bad processor but:

For most users, the Zen 4 counterparts of Zen 5 offer close to same performance for much less $$$
Not really any reason to upgrade from Zen 4 to zen 5.
Best gaming CPU's are Zen 4 7800X3D and 7950X3D

So only if you know you have software that specifically uses AVX512 and other professional uses, does it make sense to choose Zen 5 over Zen 4.

If you have a CPU that has 20%+ performance uplift, sure you can charge 30%+ for that performance premium. If you have 5-8% performance uplift don't expect people to be willing to pay more than 10-15% premium, and they would still not upgrade their CPU for such a small performance bump.

 

[Hans Gruber]

Zen 5 is not a bad processor but:

For most users, the Zen 4 counterparts of Zen 5 offer close to same performance for much less $$$
Not really any reason to upgrade from Zen 4 to zen 5.
Best gaming CPU's are Zen 4 7800X3D and 7950X3D

So only if you know you have software that specifically uses AVX512 and other professional uses, does it make sense to choose Zen 5 over Zen 4.

If you have a CPU that has 20%+ performance uplift, sure you can charge 30%+ for that performance premium. If you have 5-8% performance uplift don't expect people to be willing to pay more than 10-15% premium, and they would still not upgrade their CPU for such a small performance bump.

Get back to us with your analysis in October when Intel releases their Arrow Lake CPU's. I think we may be looking at Conroe 2.0

 

[igor_kavinski]

I think we may be looking at Conroe 2.0

Come on. That's an exaggeration. Nova Lake could be Conroe 2.0 but not Arrow Lake. It's their first (technically second since MTL-S was a dud) desktop disaggregated tile architecture. Third time's the charm so Nova Lake may be it.

 

[biostud]

Get back to us with your analysis in October when Intel releases their Arrow Lake CPU's. I think we may be looking at Conroe 2.0

You can't compare it to unreleased products of unknown performance.

Obviously, if Intel launches a much better processor at the same price point, then Zen 5 is no longer to be the best performing processor. How else would an analysis go?