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

[DisEnchantment]

Speculate at will

 

[igor_kavinski]

Linux vs. Windows results / Geekbench / Discussion Area - Primate Labs Support

support.primatelabs.com

Good answers there. Seems CPUs are able to boost higher under Linux.

 

[whoshere]

Good answers there. Seems CPUs are able to boost higher under Linux.

I've not been able to confirm that.

From .gb6:

Linux: 5059MHz (the link requires to be logged in)
Windows: 5064MHz (the link requires to be logged in)

Again, could be different compilers or compiler options.

 

[igor_kavinski]

Again, could be different compilers or compiler options.

One more theory is that there are fewer context switches happening under Linux.

There's just too much bad code running in Windows that excessively interrupts or polls the CPU.

Could even be some driver.

I'm not sure but if it's possible to run Geekbench in Safe Mode with Networking OR even better, command line Geekbench in Console Mode, I would bet that the performance would be closer to that under Linux.

 

[fastandfurious6]

like.... how hard can it be for GB to run ST on a big core automatically?

one of the most popular benchmark sw out there and still can't do the basics... in 2024

 

[Markfw]

Well, I can confirm one thing here.... Both CPUs and GPUs are allowed to work "harder" as in produce more production under linux compared to windows. This is pretty much proven in the DC forum, but you would have to read a lot to see, it, so just take my word for it.

Especially GPUs, you can see that the same GPU is using more wattage in linux as compared to windows. doing exactly the same task.

 

[StefanR5R]

If an NPU is included in the CPU like on Arrow Lake DT then the problem is solved. But on Zen5 DT it needs to be solved in some other way,

"The Problem", i.e. whether or not a given computer can be sold as a "Copilot+ PC" (vulgo "AI PC") is not a technical question, it is a marketing question which can only be answered by those at Microsoft who control this marketing program.

 

[StefanR5R]

like.... how hard can it be for GB to run ST on a big core automatically?

one of the most popular benchmark sw out there and still can't do the basics... in 2024

I am not sure if this is primarily the responsibility of the benchmark, of the user of the benchmark, or of the operating system. Arguably, the operating system's process scheduler should generally keep a single-threaded processor-intensive task on a classic core (unless the computer is switched into some sort of low-power mode maybe).

As long as we see mere database dumps of benchmark results, we are left in the dark what really happened during these GB runs. We need proper benchmark reports from reviewers who know what they are doing. At best, we are getting these after the review embargo ends, or retail availability starts. (Not that I care about GB scores personally; it's just that we have a lot of idle back and forth here in this thread about what to make of these result dumps.)

 

[igor_kavinski]

it is a marketing question which can only be answered by those at Microsoft who control this marketing program.

But they already answered it: 40 TOPS minimum!

And they will do nothing about any other NPU/GPU support for their Copilot+ stuff until they've sold a good amount of Snapdragon Elite X devices, including their own Surface laptops.

 

[MadRat]

Co-pilot wasn't too useful from playing with the GUI interface apart from potential ethical abuses. I more fear its CLI abuse by 'non-users'.

 

[igor_kavinski]

Co-pilot wasn't too useful

Here are some samples that the Copilot runtime can be used to create: https://learn.microsoft.com/en-us/windows/ai/samples/

Microsoft is depending more on the ingenuity of the developer community than any concerted in-house effort (because it seems all their experienced AI developers must've burnt out getting this Copilot runtime ready and may be on vacation, for who knows how long! ).

 

[fastandfurious6]

Here are some samples that the Copilot runtime can be used to create: https://learn.microsoft.com/en-us/windows/ai/samples/

Microsoft is depending more on the ingenuity of the developer community than any concerted in-house effort (because it seems all their experienced AI developers must've burnt out getting this Copilot runtime ready and may be on vacation, for who knows how long! ).

the first image shows editing of audio with subject: "the time I fought a bear"......

all this trouble, all these trillions of dough literally inventing solutions for fictional problems 😭😭

 

[StefanR5R]

it is a marketing question which can only be answered by those at Microsoft who control this marketing program.

But they already answered it: 40 TOPS minimum!

I don't believe that this is the full answer. ODMs/ OEMs who actually applied for the badge will know more.

 

[APU_Fusion]

This thread now has copilot discussion along with everything else. 2 👍s up

 

[fastandfurious6]

zen 6 = the zen 5 we were promised?????????? apparently coming already next year?

New AMD Zen 6 leak reveals Medusa Point for laptops alongside possible AM5 Medusa Halo APUs

While Zen 5 CPUs are yet to hit the market, Moore’s Law is Dead has already started divulging significant details regarding Zen 6. The leaker claims that the architecture could manifest in the form of “Medusa Point” for high-end notebooks, “Medusa Ridge” for desktop AM5 platform, and “Medusa...

www.notebookcheck.net

 

[igor_kavinski]

Medusa sounds a lot more menacing than Granite Ridge. Who's up for skipping Zen 5 and waiting till Zen 6?

 

[Hitman928]

zen 6 = the zen 5 we were promised?????????? apparently coming already next year?

New AMD Zen 6 leak reveals Medusa Point for laptops alongside possible AM5 Medusa Halo APUs

While Zen 5 CPUs are yet to hit the market, Moore’s Law is Dead has already started divulging significant details regarding Zen 6. The leaker claims that the architecture could manifest in the form of “Medusa Point” for high-end notebooks, “Medusa Ridge” for desktop AM5 platform, and “Medusa...

www.notebookcheck.net

Please, for the love of all that is good, let us not start a new hype train for Zen 6 based on, "MLID says. . ."

 

[Hail The Brain Slug]

Please, for the love of all that is good, let us not start a new hype train for Zen 6 based on, "MLID says. . ."

There's no way 2 CCD's, and a massive IOD with a 40CU GPU is fitting under that heatspreader. Also the narrow memory bus of AM5 would murder performance. I call total BS on this one.

 

[Jan Olšan]

One more theory is that there are fewer context switches happening under Linux.

There's just too much bad code running in Windows that excessively interrupts or polls the CPU.

Could even be some driver.

I'm not sure but if it's possible to run Geekbench in Safe Mode with Networking OR even better, command line Geekbench in Console Mode, I would bet that the performance would be closer to that under Linux.

Well, it's probably complicated. For example, some performance hit on Windows comes from VBS being enabled, if you have it. Sadly, GB reports don't show if it is. There are likely other cases where the measured performance of Windows is lower due to "features" other OSes don't have (so it means that the difference might not mean Windows bad).

I always have to think of those famous complains about Linux freezing your Desktop Environment when you copy around large files. I suspect that Linux IO and also CPU scheduling may be more tuned towards performance of running tasks and software, while Windows may be purposely aiming to actually not fully wrinch out every last CPU cycle for whatever tasks you are running to prevent those laggy desktop issues.

Basically, maximum performance/throughput × quality of service tuning.
When people see that their encoder or render takes less time, or GB6 gets better score, they may assume that Linux just is better, but what if it is at the expense of responsiveness and those frozen/laggy desktop issues? It's possible that Windows switches processes in and out more often precisely to preserve desktop responsiveness?

I used to encode with multiple encoder instances on every thread of my CPU, which means that pretty much every CPU cycle can be consumed by the encoding, and you could just work on the PC normally, with Windows (however, it does require lowering the priority of those processes). It's entirely possible that Windows scheduling policies do reflect that. So some of the perceived worse performance could just be a willing tradeoff.
Well, lots of things are like that. You can opt to run a riced Gentoo but there is a price. Windows doesn't require you to compile the software, you can download a binary of some obscure tool built in 2001 and just use it. Some people call everything cruft but quite often, it's a feature. Trading performance for features can be completely okay decision, after all, not every Linux user runs Gentoo instead of a prepackaging-based distribution.

 

[Joe NYC]

View attachment 102318

There's no way 2 CCD's, and a massive IOD with a 40CU GPU is fitting under that heatspreader. Also the narrow memory bus of AM5 would murder performance. I call total BS on this one.

It's strange that he calls a massive ~200mm SoC die with full GPU + NPU an "IOD"

 

[Det0x]

View attachment 102318

There's no way 2 CCD's, and a massive IOD with a 40CU GPU is fitting under that heatspreader. Also the narrow memory bus of AM5 would murder performance. I call total BS on this one.

There is quite a bit more room under the IHS to fit a bigger IOD if you move things around.. Maybe 2.3x the current IOD size by eyeballing it (?)

This is one of my delidded 7950X
Red square is the IHS ~size
Blue square is the ~size of a 7000 series IOD

Zen4 IOD is 124.7 mm² according to techpowerup

125mm2 * 2.3 = 287mm2

 

[igor_kavinski]

That doesn't look good for Zen 5: https://www.guru3d.com/story/intel-...thread-performance-improvement-over-i913900k/

 

[Det0x]

That doesn't look good for Zen 5: https://www.guru3d.com/story/intel-...thread-performance-improvement-over-i913900k/

Already confirmed fake and discussed in the other thread, not much more to say
("20A node" also gave it away)

不是，哥们，造谣真没意思【高通吧】_百度贴吧

ä¸æ¯ï¼å¥ä»¬ï¼é è°£ç没..iç²å®äºæ±æ¯çä¸ææ§ææiå§å·²ç»åè¿ä¸æ¬¡è¾è°£è´´äºï¼åç»éå§å¤§ä¼æ±æ¥ä¸ä¸ç®åarrow lakeçesè¿å±é¦åintelçesçï¼æç§è¿åº¦ä¼ååæä¸ç§es1:åçesï¼é常é¢çæä½ä¸ä¼´éçå¾å¤bugï¼ä¸

tieba.baidu.com

 

[igor_kavinski]

Already confirmed fake and discussed in the other thread, not much more to say

Thanks! Been hard to follow that thread. I have over 30 pages I need to skim through :S

 

[yuri69]

I always have to think of those famous complains about Linux freezing your Desktop Environment when you copy around large files. I suspect that Linux IO and also CPU scheduling may be more tuned towards performance of running tasks and software, while Windows may be purposely aiming to actually not fully wrinch out every last CPU cycle for whatever tasks you are running to prevent those laggy desktop issues.

Can you point me towards a guide how to replicate the Linux IO issues? Personally, I have never encountered such issue. It feels like an implementation issue of Files/Dolphin/Nautils.

Well, lots of things are like that. You can opt to run a riced Gentoo but there is a price. Windows doesn't require you to compile the software, you can download a binary of some obscure tool built in 2001 and just use it. Some people call everything cruft but quite often, it's a feature. Trading performance for features can be completely okay decision, after all, not every Linux user runs Gentoo instead of a prepackaging-based distribution.

There are static binaries that you can run old binaries on Linux too. No need to recompile.

You don't have to run a source-based distro to get perf gains. Maintainers are not dumb. Mainstream distros are shifting towards AVX2 aka level v3...

 

[Cllaymenn]

I have a different table for DDR5-8000. It doesn't matter. The core has potential. Whether everything can work out in the end is the big question.

Can you arrange such a test in AIDA64? It is more important than FP32/64 and AES
This is my result with 5950X without OC but with good optimization