Further proof that Windows 10 is causing a perf hit with the Ryzen CPU

[Taf the Ghost]

55 minutes ago, LAwLz said:

How big was the improvement in terms of percentage points?

No clue.  Since it's new tech, Linux folks tend to spend more time making sure it works, firstly, though since Ryzen seems a lot more like the first SKU in their server-chips more than anything else, I imagine they had the scheduler ready for release to the distros early.

 

The main take away from the early Linux benchmarks is that Ryzen is a monster compile/render chip.  The "Naples" server ones are going to be super-hot commodities, which kind of explains their release cycle.  R7s will appeal to the Bleeding Edge users, which will create buzz (and benchmarks) for the corporate buyers.

[cj09beira]

50 minutes ago, MageTank said:

I wouldn't be so sure. Knights Landing and Skylake-E (AVX-512) are coming. Intel has plenty of tricks up their sleeves when it comes to the HPC market, and are willing to flex those muscles if AMD gets too close with Naples. I won't say AMD doesn't stand a chance, simply because I do not know yet, but I certainly wouldn't say they are going to "murder Intel's server business" because that's simply not going to happen in a single generation.

Plus we know ryzens avx are only 128bit

[jagdtigger]

I wouldnt be surprised if it turns out that MS only optimized for intel chips... I relly hope that AMD can keep up the good work, now the only work left is for the software devs to start making proper multi threaded applications.

Edited March 8, 2017 by wkdpaul
Cleaned up

[cj09beira]

50 minutes ago, MageTank said:

I wouldn't be so sure. Knights Landing and Skylake-E (AVX-512) are coming. Intel has plenty of tricks up their sleeves when it comes to the HPC market, and are willing to flex those muscles if AMD gets too close with Naples. I won't say AMD doesn't stand a chance, simply because I do not know yet, but I certainly wouldn't say they are going to "murder Intel's server business" because that's simply not going to happen in a single generation.

Plus we know ryzens avx are only 128bit

[XenosTech]

Well AMD's design is way different from Intel's so if the scheduler is treating AMD's chip like Intel's that'll lead to problems with SMT

Edited March 8, 2017 by wkdpaul
cleaned up

[cj09beira]

45 minutes ago, XenosTech said:

 

Well AMD's design is way different from Intel's so if the scheduler is treating AMD's chip like Intel's that'll lead to problems with SMT

Its bad also for when 2 threads share info between the two if they are on diferent ccxs then it has to go to main memory. 

Edited March 8, 2017 by wkdpaul
cleaned up

[Morgan MLGman]

46 minutes ago, XenosTech said:

Well AMD's design is way different from Intel's so if the scheduler is treating AMD's chip like Intel's that'll lead to problems with SMT

 

25 minutes ago, cj09beira said:

Its bad also for when 2 threads share info between the two if they are on diferent ccxs then it has to go to main memory. 

There are obviously lots of things to consider and fix, but I suggest you take a look at this:

https://www.techpowerup.com/231268/amds-ryzen-cache-analyzed-improvements-improveable-ccx-compromises

 

Overall, there's actually a lot of things to be improved, whether it's the scheduler in W10 or its power management as there are very mixed results in reviews out there, people had trouble maintaining clock speeds in their R7 CPUs under Win10, that's most likely a cause of power management settings and the HPET feature being not configured properly for Ryzen.

 

It's just that people need to understand that such a brand-new architecture will have more or fewer issues at launch, but it surely will. Intel hasn't done a completely redesigned architecture for a very long time and Kaby Lake is just a laugh for me. Of course it's good, but it's basically just as good as Skylake. Intel branded their Kaby lake as 15% faster than Skylake, yeah, "faster" my ass with higher stock clock speeds and a bit higher OCing potential while leaving IPC and the technological process the same... So overall, people need to sit on their asses and wait for the promised fix.

[WkdPaul]

Thread cleaned, please keep the Community Standards in mind.

 

It's not allowed to attack others or to talk about moderation (ie: bans and warnings).

[XenosTech]

2 hours ago, cj09beira said:

Its bad also for when 2 threads share info between the two if they are on diferent ccxs then it has to go to main memory. 

Or the threads are treated as physical cores

 

1 hour ago, Morgan MLGman said:

 

There are obviously lots of things to consider and fix, but I suggest you take a look at this:

https://www.techpowerup.com/231268/amds-ryzen-cache-analyzed-improvements-improveable-ccx-compromises

 

Overall, there's actually a lot of things to be improved, whether it's the scheduler in W10 or its power management as there are very mixed results in reviews out there, people had trouble maintaining clock speeds in their R7 CPUs under Win10, that's most likely a cause of power management settings and the HPET feature being not configured properly for Ryzen.

 

It's just that people need to understand that such a brand-new architecture will have more or fewer issues at launch, but it surely will. Intel hasn't done a completely redesigned architecture for a very long time and Kaby Lake is just a laugh for me. Of course it's good, but it's basically just as good as Skylake. Intel branded their Kaby lake as 15% faster than Skylake, yeah, "faster" my ass with higher stock clock speeds and a bit higher OCing potential while leaving IPC and the technological process the same... So overall, people need to sit on their asses and wait for the promised fix.

Well according to he who shall not be named, the task scheduler in windows doesn't need to be optimised for AMD since multi-core cpu's have been around for years and AMD did a shit job... That would explain why the R7's shows better performance under win7 and Linux after their scheduler was updated to work with ryzen

[MageTank]

9 hours ago, cj09beira said:

Plus we know ryzens avx are only 128bit

Are you sure? Last I checked, it was both 128 and 256. 8 x 128-bit op/clock or 8 x 256-bit op/clock. Haswell on the other hand, can do 16 x 128-bit op/clock or 8 x 256-bit op/clock. Meaning Ryzen's AVX1 (AVX-128) is half as good as Haswells, but it's AVX2 (AVX-256) is on par. 

[cj09beira]

4 minutes ago, MageTank said:

Are you sure? Last I checked, it was both 128 and 256. 8 x 128-bit op/clock or 8 x 256-bit op/clock. Haswell on the other hand, can do 16 x 128-bit op/clock or 8 x 256-bit op/clock. Meaning Ryzen's AVX1 (AVX-128) is half as good as Haswells, but it's AVX2 (AVX-256) is on par. 

i read wrong then, i got the idea that they had 128 and they had 256 at half speed (emulating it ).

good to know they are on par on avx 256