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

[DisEnchantment]

Speculate at will

 

[Bigos]

A lot of software has components that can run in parallel independently and ones that cannot. This can be usually seen as "a wall" in scaling, where such software no longer scales with more cores/threads.

One example would be code compilation where you can split the work in many processes but eventually have to link everything together. Even though linking was historically a serial process, it has been made parallel to some degree so it is not a complete single-thread part. Still, when we compare 64+ core CPUs one can see they do not scale much anymore (depends on source code size).

What am I trying to get to? There are a *a lot* various MT behaviors. Scaling varies from software to software, it also depends on the data. Trying to model one of them and telling "only this one represents client workloads" makes the benchmark pretty useless to my eyes.

Speaking of code compilation, gb6 has a subtest that compiles 96 files in parallel. That is not a lot of parallelism compared to for example compiling the Linux kernel or Chromium or LLVM itself. This subtest then indeed appears to work independently and not on a shared data, but the flaw is that it is pretty small. And the long compile times are usually complained when working on big source code, reducing the time there is more important than making the 96 files case better (of course there is a parallelism limit here as well, depending on the source code).

 

[qmech]

You only need to read page 2 that contains the summary, and then click on Multithreading for example. It's a well structured document.

Here's the text:


And here's a shorter version in case 3 paragraphs is too much, emphasis mine:

There's a thread for discussing Geekbench, so there's probably a limit to how many pages we should devote to it here, so in an effort to be succinct, the issue with GB6 multi-core is this:

It attempts to score the performance of a single multithreaded task run all by its lonesome. This is a very good approximation of how current phones work, but not how many people use their laptops/desktops, especially if you multitask.

This is a deliberate design choice, and a completely fair one given how GB is used today, as is the decision to not include any game-like components (in either single or multi-core). Again, a completely valid design choice, since there are many game-specific benchmarks, but it does underscore the reality of what GB6 multi-core measures (and this is stated in the doc you linked, although sadly without specifics): The performance of a single, multithreaded app run alone.

Finally, the way they sized their datasets, unfortunately, results in large caches being of little use while bandwidth to main memory retains significant importance. This rarely reflects actual performance in user applications and if GB were open about their internal set of apps that the benchmark is to mimic, we would likely see that to be the case there as well.

 

[coercitiv]

I agree that people with strong sentiments about GB viability as a consumer MT benchmark should probably find and use the GB thread.

Going back to Zen 5, have there been any recent hints as to when we might expect the first consumer products with the new arch?

 

[tamz_msc]

This is a very good approximation of how current phones work, but not how many people use their laptops/desktops, especially if you multitask.

This assumption is basically the result of a near-universal set of beliefs that all culminate in the following -

"A CPU benchmark is not a good one unless it behaves the same way as Cinebench(before the 2024 edition)/Blender/SPEC_rate"

There is nothing that makes this belief firmly rooted in terms of software behavior on multi-processor environments. As it is, it's basically an argument of faith - "believe me bro - this is a great CPU because it scores xx % better in Cinebench than the competition".

Let's take multitasking - the assumption in case of phones is that people do at most one thing at a time that involves interaction with the display, while all other tasks are relegated to the background. Well, is it any different from how we use PCs?

Multitasking isn't "let's render something on blender, while baking lighting to the Unreal Engine game, and let's run both on the CPU!" - That's plain stupid, and will result in a PC that is unusable for anything else.

Multitasking is always about offloading different tasks to different specialized hardware - so compiling code = CPU, rendering 3D = GPU, watching video = iGPU video engine, etc.

 

[Thunder 57]

A lot of software has components that can run in parallel independently and ones that cannot. This can be usually seen as "a wall" in scaling, where such software no longer scales with more cores/threads.

One example would be code compilation where you can split the work in many processes but eventually have to link everything together. Even though linking was historically a serial process, it has been made parallel to some degree so it is not a complete single-thread part. Still, when we compare 64+ core CPUs one can see they do not scale much anymore (depends on source code size).

What am I trying to get to? There are a *a lot* various MT behaviors. Scaling varies from software to software, it also depends on the data. Trying to model one of them and telling "only this one represents client workloads" makes the benchmark pretty useless to my eyes.

Speaking of code compilation, gb6 has a subtest that compiles 96 files in parallel. That is not a lot of parallelism compared to for example compiling the Linux kernel or Chromium or LLVM itself. This subtest then indeed appears to work independently and not on a shared data, but the flaw is that it is pretty small. And the long compile times are usually complained when working on big source code, reducing the time there is more important than making the 96 files case better (of course there is a parallelism limit here as well, depending on the source code).

Soundles like you've read up on Amdahl's law. Sounds like you hit the nail on the head in that last paragrapgh.
.

I agree that people with strong sentiments about GB viability as a consumer MT benchmark should probably find and use the GB thread.

Going back to Zen 5, have there been any recent hints as to when we might expect the first consumer products with the new arch?

Ask our all things Zen overlord @adroc_thurston. This was discussed breifly recently. I said March-April, but I know nothing. He replied

~April'24 or therein.

. So hey, maybe we aren't all that far off!

 

[TESKATLIPOKA]

Ask our all things Zen overlord @adroc_thurston. This was discussed breifly recently. I said March-April, but I know nothing. He replied

. So hey, maybe we aren't all that far off!

That's probably desktop one, mobile will be later.

 

[tamz_msc]

Trying to model one of them and telling "only this one represents client workloads" makes the benchmark pretty useless to my eyes.

That's settled it then - Cinebench/Blender/SPEC_rate are all useless. Only Geekbench 6 is valid.

 

[tamz_msc]

Obligatory, to those concerned people:

 

[moinmoin]

@tamz_msc Please head here: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...and-calibrated-against-core-i7-12700.2610597/

 

[eek2121]

A lot of software has components that can run in parallel independently and ones that cannot. This can be usually seen as "a wall" in scaling, where such software no longer scales with more cores/threads.

One example would be code compilation where you can split the work in many processes but eventually have to link everything together. Even though linking was historically a serial process, it has been made parallel to some degree so it is not a complete single-thread part. Still, when we compare 64+ core CPUs one can see they do not scale much anymore (depends on source code size).

What am I trying to get to? There are a *a lot* various MT behaviors. Scaling varies from software to software, it also depends on the data. Trying to model one of them and telling "only this one represents client workloads" makes the benchmark pretty useless to my eyes.

Speaking of code compilation, gb6 has a subtest that compiles 96 files in parallel. That is not a lot of parallelism compared to for example compiling the Linux kernel or Chromium or LLVM itself. This subtest then indeed appears to work independently and not on a shared data, but the flaw is that it is pretty small. And the long compile times are usually complained when working on big source code, reducing the time there is more important than making the 96 files case better (of course there is a parallelism limit here as well, depending on the source code).

Yes, however, the issue is that they are benchmarking the efficiency of the software rather than testing the performance potential of the CPU. That is the beef I have with Geekbench 6. I don’t care how poorly threaded software can be. I care about what the CPU is capable of.

I can (and have) write a highly parallel piece of software. Software is constantly changing/improving. More cores will eventually get utilized as core counts increase. Some software today can already take advantage of highly parallel workloads. Not showing peak performance capability is a mistake.

 

[IEC]

@tamz_msc Please head here: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...and-calibrated-against-core-i7-12700.2610597/

Or watch the latest GN video and despair at how bad the Intel parts are for power efficiency:


Zen 5 barely has to iterate to dominate. Intel needs to step it up.

 

[Thunder 57]

Or watch the latest GN video and despair at how bad the Intel parts are for power efficiency:
View attachment 90738

Zen 5 barely has to iterate to dominate. Intel needs to step it up.

Just watched that. It was a bloodbath.

 

[soresu]

Or watch the latest GN video and despair at how bad the Intel parts are for power efficiency:
View attachment 90738

Zen 5 barely has to iterate to dominate. Intel needs to step it up.

Is that Meteor Lake?

 

[Thunder 57]

Is that Meteor Lake?

No it was a comparison of desktop chips. Someone bet that a 14700k was more efficient than a 7800X3D. It didn't go well.

 

[tamz_msc]

Or watch the latest GN video and despair at how bad the Intel parts are for power efficiency:
View attachment 90738

Zen 5 barely has to iterate to dominate. Intel needs to step it up.

Ah yes, a normal adult with things to do plays Cyberpunk for 8 hours everyday.

 

[Geddagod]

Ah yes, a normal adult with things to do plays Cyberpunk for 8 hours everyday.

I'm genuinely shocked at how some people manage to do that. Even at the peak of my "yolo" and "fck it I don't have any responsibilities" (aka senior year of HS) phase, I only managed to play valorant like 6 hours a day max, spread out between the afternoon till like 3 AM, and on the weekends, while grinding to Diamond.

No it was a comparison of desktop chips. Someone bet that a 14700k was more efficient than a 7800X3D. It didn't go well.

It was very weird and quirky but idk why someone who has as much "journalistic integrity" as Steve from Gamers Nexus can't just present data for the sake of presenting data. It's also strange how defensive and "extra" these tech youtubers are on twitter (not just GN) ¯\_(ツ)_/¯

 

[gdansk]

14900K users channeling Jensen in reverse. The less you game, the more you save.

 

[Thunder 57]

I'm genuinely shocked at how some people manage to do that. Even at the peak of my "yolo" and "fck it I don't have any responsibilities" (aka senior year of HS) phase, I only managed to play valorant like 6 hours a day max, spread out between the afternoon till like 3 AM, and on the weekends, while grinding to Diamond.

It was very weird and quirky but idk why someone who has as much "journalistic integrity" as Steve from Gamers Nexus can't just present data for the sake of presenting data. It's also strange how defensive and "extra" these tech youtubers are on twitter (not just GN) ¯\_(ツ)_/¯

And I remember being in high school certainly not playing games for 8 hours a day but still warming up my room about 10 degrees compared to others. It was annoying.The extra cost? Meh. The extra heat, unbearable.

I also don't see what is so "defensive" about disputing a nonsense claim. Personaly I like GN. Even moreso after the LTT debacle.

 

[Geddagod]

And I remember being in high school certainly not playing games for 8 hours a day but still warming up my room about 10 degrees compared to others. It was annoying.The extra cost? Meh. The extra heat, unbearable.

I live in the southern part of the US and game on a 12900H, the extra heat part is deff very relatable in the already pretty hot summers lol

I also don't see what is so "defensive" about disputing a nonsense claim. Personaly I like GN. Even moreso after the LTT debacle.

It's petty. Disputing a nonsense claim is fine, but like bringing it up multiple times in a video "in response" to a troll post is a bit much. I mean it's good entertainment, but like... srsly?
But perhaps the most egregious example of stuff like this is just like most of HWUB's entire twitter interactions. Responding to trolls is not something that exactly helps their image, IMO.

 

[igor_kavinski]

I actually prefer Intel being dragged in the mud over their power consumption. It should help to motivate them

 

[Thunder 57]

I live in the southern part of the US and game on a 12900H, the extra heat part is deff very relatable in the already pretty hot summers lol

It's petty. Disputing a nonsense claim is fine, but like bringing it up multiple times in a video "in response" to a troll post is a bit much. I mean it's good entertainment, but like... srsly?
But perhaps the most egregious example of stuff like this is just like most of HWUB's entire twitter interactions. Responding to trolls is not something that exactly helps their image, IMO.

I am in the south as well these days, so I value low TDP. Even when I lived further north, the extra heat was a nuisance. In the Winter it wouldn't be bad, but I could always sleep with warmer blankets. During the summer it just sucked! Back in the days before throttling and power gating and everything just ran all out.

As far as GN and others, I'm sure they do it for the money, much like LTT. It's a bit sad that is the way these days. I much prefer written articles like Anand and Tom's of years old did. I welcome GN's new written reviews though they are infrequent and after video reviews which are surely more profitable.

 

[Thunder 57]

I actually prefer Intel being dragged in the mud over their power consumption. It should help to motivate them

Maybe they'll have a chip called Ice Lake that sips power soon! /s

 

[MrTeal]

Maybe they'll have a chip called Ice Lake that sips power soon! /s

You misunderstand. Ice Lake is designed to take a glacier or ice cap and terraform it into a lake.

 

[Kepler_L2]

I actually prefer Intel being dragged in the mud over their power consumption. It should help to motivate them

Won't help when this is what Intel thinks (yes this is from an official Intel slide)

 

[inf64]

Won't help when this is what Intel thinks (yes this is from an official Intel slide)View attachment 90771

This is now pure desperation from intel's side. From AMD being "left out in a rear view mirror", to these pathetic performance and perf./watt numbers from "game changing" MTL architecture, they are losing the plot.