Official AMD Ryzen Benchmarks, Reviews, Prices, and Discussion

[lolfail9001]

Please: no Tweaktown links until Tweaktown fixes their broken ad system.

Uh, sorry, i forgot about the ads existing on internet.

On a side note, My 960 Evo 256GB NVMe performance on Ryzen is perfectly in line with expectation in AS SSD... as in, it's faster than on Intel.

Dayum, what is up with 4k writes on 2680v4. I mean, hell Toshiba SSD in random Dell laptop has better 4k writes as result. Granted, Ryzen results directly match TT's.

 

[cytg111]

http://www.tweaktown.com/articles/8073/amd-ryzen-ssd-storage-performance-preview/index3.html

Look at low block size write results.

Why isnt this a corner case of a corner case and something fixable with a firmware/sw update?

 

[sirmo]

That legitreviews article is a little sad. Basically, it's been so long since AMD had a new socket and CPU someone actually wants to buy they didn't know what to do.

On the other hand, it's an improvement. From AMD screws up a roll out of a mediocre product to AMD screws up a roll out of a great product.

Baby steps, AMD!

I don't buy that legitreviews article at all. For multiple reasons. Lisa Su has been saying beginning of the year (2017) for market wide availability since like a Q1 of 2016. And it never really changed. If anything it was pushed back some, because she mentioned that we might see limited availability at the end of 2016. So I don't believe motherboard manufacturers didn't see it coming. The whole world knew it was coming. Also AMD is throwing motherboard manufacturers a bone here. X370 motherboards are cheaper to make (cheaper socket, cheaper chipset) than Z270 motherboards, yet they are charging about the same for them.

Now it is possible BIOS manufacturers wanted more time, or didn't get as good a support, I buy that. But to not know it was coming and not know that there was a big pent up demand, is BS. We all knew it.

Also AMD only launched the top Ryzen 7 parts. Ryzen 5 isn't even here and that's where the real volume is.

The article reeks of made up BS.

 

[unseenmorbidity]

Hey have any of you guys noticed this before?

CPUs supported by Biostar motherboard: Ryzen R7, Athlon™ II X4, & Bristol A6,A8,A10,A12

 

[piesquared]

What are the chances that AMD will start branding memory again?

 

[lolfail9001]

Why isnt this a corner case of a corner case and something fixable with a firmware/sw update?

Because my hypothesis is that this has the same nature as games performance being... worse than it should be, that's the wording, and terrible drawcall performance in that D3D9 synthetic demo: syscall performance.
And if it is, then no amount of firmware or software updates will fix it. Yes, SMT and CCX shuffling play their role, but they are not going to explain StarCraft 2 results, are they? Granted, 5960X does pretty terribly in it too.

CPUs supported by Biostar motherboard: Ryzen R7, Athlon™ II X4, & Bristol A6,A8,A10,A12

Bristol Ridge has an Athlon X4 in it, what's your point?

Though you got me thinking: AMD promised us 82 motherboards. Where is another 40 or so?

 

[PotatoWithEarsOnSide]

The cheaper A320s aren't out yet.

 

[nathanddrews]

This guy's channel is comparing Windows 7 to Windows 10 performance to work out the scheduler bug. CS:GO sees a big boost here, but he's going to rerun the game using the benchmark. Someone already posted the Mankind Divided benchmark. Hopefully this channel posts more than average fps data.

 

[Agent-47]

I should also note gameplay was smooth and perfect the entire time - I was surprised to see that these dips existed. I also have frametimes that align with this, but graphing that is a bit more tricky as I need to time-align them (which I can do from file creation and modification times).

For all the BF lovers! Too bad he did not have access to a i7. or a newer generation of Core.

but its interesting to see core scaling from 4c8t to 16t in ryzen

 

[looncraz]

Okay, so I've done about 7 hours of bench-marking on Ryzen on Windows 10 build 1511 (with no updates). Then I went to the pre-release fast ring and am now running the bleeding edge build 1607 with all updates to see if Microsoft has done anything worthwhile (hint: they've done exactly nothing).

I can't disable SMT in the ASRock BIOS - or I surely can't find it, but I've also partly tested 4+0 core configuration.

The Windows 10 scheduler is royally fscked. It, indeed, does not schedule with priority on real cores. But, worse than that, it avoids the second half of all logical CPUIs as if they were SMT cores (with balanced power setting).

Now, I set thread affinity on Cinebench after it began to run, with a fixed 3GHz clock speed and high-performance power settings. The logical action is to disable every other logical core in the affinity mask, which I did, enabling 0, 2, 4, & 6.

This was the, very confusing, result:

The Windows 10 scheduler forced ALL threads onto just two cores! I didn't let this finish as multiple threads were not even being scheduled during this time and Cinebench showed no progress for three or four threads for a good minute (yes, this ran forever...).

Only setting affinity to cores 0,1,2,3 did what was expected (that's dual core, however) - the entire second half of the CCX remained nearly completely dormant and a third core came online to handle ancillary background processing.

On the good news front, 3.425GHz Ryzen 1700X eliminates the bottlenecks my 2600k@4.5Ghz has:

Also, I haven't touched memory - it's running at DDR4-2133 CL15-15-15 2T (yes, 2T!).

What is really strange is that sometimes Windows 10 does the right thing - such as in Unigine benchmarks.

I am going to run some BF1, image this drive, then try Windows 7 tomorrow.

For now, you can see all my relevant screenshots here:

http://files.looncraz.net/zen/Win10/

 

[lolfail9001]

The Windows 10 scheduler is royally fscked. It, indeed, does not schedule with priority on real cores. But, worse than that, it avoids the second half of all logical CPUIs as if they were SMT cores.

Wait, what. What dark sorcery did MS pull with it's scheduler that it confuses first bit as SMT for last?

it's running at DDR4-2133 CL15-15-15 2T (yes, 2T!).

Wait, what!? That's the first i hear of CR2 on Ryzen.

 

[looncraz]

Wait, what. What dark sorcery did MS pull with it's scheduler that it confuses first bit as SMT for last?

It's not even reliably doing it.

Unigine Heaven: SMT logical cores all sleeping, real cores used:

Battlefield 4:

These are both in high performance mode, naturally.

Interestingly, I'm also not seeing load balancing going on at all with Ryzen reliably - even within a CCX with the other CCX parked... but it sometimes does it... then doesn't...

Wait, what!? That's the first i hear of CR2 on Ryzen.

Yup, that's just what it set it to by default. I didn't find anywhere I could change it in the BIOS.

 

[Mockingbird]

AMD Ryzen and the Windows 10 Scheduler - No Silver Bullet

https://www.pcper.com/reviews/Processors/AMD-Ryzen-and-Windows-10-Scheduler-No-Silver-Bullet

 

[looncraz]

AMD Ryzen and the Windows 10 Scheduler - No Silver Bullet

https://www.pcper.com/reviews/Processors/AMD-Ryzen-and-Windows-10-Scheduler-No-Silver-Bullet

I responded to that right here:

https://forums.anandtech.com/threads/ryzen-strictly-technical.2500572/page-19#post-38788377

In short - Yes, there's a problem. And it only happens with Ryzen.

 

[IEC]

More tweaking to come tomorrow. I've managed 4.11GHz all core stable for now @ 1.395V

http://valid.x86.fr/pzery0

 

[Dygaza]

Looncraz, could you possibly run some benchmarks with custom affinities where you bind threads to different cores. You can do that easily with process hacker.

I'm especially interested in seeing perf uplift when you bind game mainthread and driver thread to one ccx.

 

[looncraz]

Looncraz, could you possibly run some benchmarks with custom affinities where you bind threads to different cores. You can do that easily with process hacker.

That is some of what I've been posting . It doesn't work properly in Windows 10 on Ryzen (at least not with Cinebench R15 - literally the only app with which I tried this ) (to be clear - it works perfectly fine with the same version of Windows 10 and am i7 2600k).

I am spending more time on it tomorrow and will try it with more apps.

The default Windows 10 power settings allows dual core turbo and XFR, but if you set "high performance" you are limited to the all-core turbo and XFR. So only 3.5GHz **MAX** on a 1700X.

The fact that so many people have been showing higher performance with what likely amounts to lower clock speeds in games is one of the dead giveaways that Windows 10 is genuinely broken when handling Ryzen.

 

[piesquared]

Has anyone tried G.Skill's new Ryzen FlareX modules?

 

[lolfail9001]

That is some of what I've been posting . It doesn't work properly in Windows 10 on Ryzen (at least not with Cinebench R15 - literally the only app with which I tried this ) (to be clear - it works perfectly fine with the same version of Windows 10 and am i7 2600k).

I am spending more time on it tomorrow and will try it with more apps.

The default Windows 10 power settings allows dual core turbo and XFR, but if you set "high performance" you are limited to the all-core turbo and XFR. So only 3.5GHz **MAX** on a 1700X.

The fact that so many people have been showing higher performance with what likely amounts to lower clock speeds in games is one of the dead giveaways that Windows 10 is genuinely broken when handling Ryzen.

Did you try that without disabling cores? I mean, it probably is broken, but i am looking for validation of pcper findings if there is any.

 

[Dygaza]

That is some of what I've been posting . It doesn't work properly in Windows 10 on Ryzen (at least not with Cinebench R15 - literally the only app with which I tried this ) (to be clear - it works perfectly fine with the same version of Windows 10 and am i7 2600k).

I am spending more time on it tomorrow and will try it with more apps.

The default Windows 10 power settings allows dual core turbo and XFR, but if you set "high performance" you are limited to the all-core turbo and XFR. So only 3.5GHz **MAX** on a 1700X.

The fact that so many people have been showing higher performance with what likely amounts to lower clock speeds in games is one of the dead giveaways that Windows 10 is genuinely broken when handling Ryzen.

Are you really setting affinity for single threads, or simply just forcing whole application to run on those cores. I'm talking about binning specific gamethreads for each cores separately. If you just set affinity for application, win scheduler can still bounce them around ccx'es.

 

[krumme]

And yet there are still some people on here insisting that there is something wrong with anyone who would buy an Intel system right now.

I got my asus b350m up and running 3 hrs before launch and in the Ryzen builders thread straight out said the software was crap, buggy, hang in bios update and was generally unfinished and the bios was extremely beta and buggy - before the reviews hit. So i would say people got it straight.

Now all those problems will be fixed within a few months and some will get a small boost from better memory performance due to new bios. But zen will still have a very high latency. And Surely windows scheduler will also be fixed. In the simples form of fix we pretty much know the results and they vary from 0 to 20% improvement in games. Still Intel will have far the faster gaming solution due to latency advantage. No doubt.

What can then happen is if ms gives us a sceduler that is directly programmed for the 2xccx in zen and not only a 2 processor simple fix taken from windows 7. I think its safe to say they will do so to remain a competitive os vs eg Linux.

What is more unsecure is to what degree the game developers will adapt and program for zen. If amd want to get close to Intel that needs to be done. If. Then in those games the difference will imo be marginal and the more core advantage for zen could tilt even for gamers looking with a shorter 3 years horizon.

Its also vital for zen plus that it happens because with a fix of latency and a small other places it looks to be more or less the same stuff like an Intel processor for gaming. But it demands windows fix and game dev adopting.

Btw hopefully 1700 non x will fall to 280usd during fall and the cheapest 350 boards to 70usd. As it is 330 plus 90 is imo straight out enthusiast and prevent a lot of people upgrading.

There is a motherboard shortage for now so people might as well wait a month and see how things pan out.

 

[looncraz]

Are you really setting affinity for single threads, or simply just forcing whole application to run on those cores. I'm talking about binning specific gamethreads for each cores separately. If you just set affinity for application, win scheduler can still bounce them around ccx'es.

I haven't find an application that lets you do that... including Process Lasso.

You are limited to process affinity, AFAIK. Usually that is all you need, anyway. You might gain 1~2% in some cases with sticking certain threads to a specific core, but I'm not actually seeing Windows 10 doing much load balancing on Ryzen - so, sometimes, the threads, contrary to my earlier belief, are not bouncing between CCXes nor are they bouncing from core to core... but, sometimes, they are...

Cinebench R15, 3Ghz fixed, 4 threads (no affinity), 4C/8T:

... As you might imagine, the results of this configuration are not as much better than four threads running on two cores as they should be... because it's only running on three cores - and is, oddly, load-balancing two threads between four logical CPUs...

When using all eight cores, however, with default settings, threads often reside entirely on one CCX and the scheduling algorithms seem to treat each logical core as if it were a real core, rather than jumping to the other CCX in a manner that would benefit performance. The other CCX will remain parked and not come to life until there are a LOT of heavy threads... and there is a 30ms or so delay before that happens... that is enough for several frames of a game to pass.

This is Battlefield 4 with default settings:

BF4 isn't being CPU bottlenecked, either, so this is actually quite acceptable behavior... in this one case, anyway... and it actually ran at a good clock speed.... which is another tricky bit...

Here is the system running Unigine Heaven with default settings... it clocks down to 2.45GHz. It was not even slightly warm, Windows was simply setting P-states and slowing down the frequency - and it parked all but two cores. I would not be surprised to see this happen in quite a few canned gaming benchmarks (though not necessarily to this degree).

 

[Dygaza]

I haven't find an application that lets you do that... including Process Lasso.

I already mentioned you can do it with process hacker.
http://processhacker.sourceforge.net

Go to properties of application, and find threads tab. Or simply press threads button on top menu after selecting application. Setting affinity works same way as setting them for apps.

 

[lolfail9001]

Here is the system running Unigine Heaven with default settings... it clocks down to 2.45GHz. It was not even slightly warm, Windows was simply setting P-states and slowing down the frequency - and it parked all but two cores. I would not be surprised to see this happen in quite a few canned gaming benchmarks (though not necessarily to this degree).

I mean, if it was Heaven with default settings, it was probably not stressing CPU anyways, so kind of different from usual canned benchmarks that appeared in reviews.

When using all eight cores, however, with default settings, threads often reside entirely on one CCX and the scheduling algorithms seem to treat each logical core as if it were a real core, rather than jumping to the other CCX in a manner that would benefit performance. The other CCX will remain parked and not come to life until there are a LOT of heavy threads... and there is a 30ms or so delay before that happens... that is enough for several frames of a game to pass.

Wait, do you have any screens of CB on 8 core config, affinity or not? Just as reference point, because i did not see it in your array of screens.

 

[looncraz]

I already mentioned you can do it with process hacker.
http://processhacker.sourceforge.net

Go to properties of application, and find threads tab. Or simply press threads button on top menu after selecting application. Setting affinity works same way as setting them for apps.

I see... will have to play with that then! I thought you said Process Lasso, which doesn't work.

Setting affinity on Windows 10 and Ryzen, period, doesn't seem to work properly for me. Bed time, ish, will play tomorrow!

I mean, if it was Heaven with default settings, it was probably not stressing CPU anyways, so kind of different from usual canned benchmarks that appeared in reviews.

Wait, do you have any screens of CB on 8 core config, affinity or not? Just as reference point, because i did not see it in your array of screens.

I always max everything during testing - still, it is a low load - but enough of a load that my 2600k on Windows 10 will run steady at 4.5Ghz, whereas Ryzen will flirt with idle frequencies.

As for CB on 8 cores and 16 threads - do you mean with affinity set? Because it's like this:

That is 0, 2, 4, 6, 8, 10, 12, 14 affinity (avoid logical cores in Process Lasso).

I replicated this with every single application. I've also verified that Windows does not do this with the 2600k.

Of you mean without thread affinity... imagine 100% CPU usage for all cores for a blip, because it's done before you can sneeze.