AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

[Phynaz]

I originally did this vs the Ryzen baseline when it showed up, but decided to run some extra numbers.

Snip

Props dude, really nice!

 

[dfk7677]

I would say the Physics benchmark gets bottlenecked by dual DDR4-2400 around the 750 mark.

Looking at the algorithm used in the Prime benchmark it is Integer and branching (I think). If we take into account that Ryzen is >20% better clock for clock in Int, that leaves us to either bad branching or bad SMT in Prime benchmark (or both).

 

[Agent-47]

Interesting results here! The HT effect might be related to cache thrashing, partitioned core resources, fighting for sparse resources (which incl. cache and mem channels of course). Just fighting for an IDIV unit alone (if used at all) wouldn't cause such a significant drop.

Someone wondered about the string sorting. That test's lower mem dependency could come from cache blocking or easier to detect access patterns for prefetching, or both.

OTOH Passmark has a history of fiddling with IDIV. It's past CPU benchmark result measured the runtimes of loops with special instructions, one being IDIV. But it ran the same amount of instructions (IIRC) and created the score from the whole runtime. Now imagine 1 IDIV loop contributing maybe 80% of a benchmarks runtime if there is no HW divider. This way the per patch deactivated Llano HW divider got some attention.

http://www.passmark.com/forum/performancetest/3705-amd-llano-a-series-benchmark-and-cpu-bug

Very interesting read this thread! Thank you!

While the thread confirms that PN score is IDIV based, but should the Integer results not have poor results too? At least this thread seems to suggest it. And if it an error on AMD 12h CPUs, should it now have been fixed by now?

But I agree with you on the string sorting and the fight for cache. I also think it might be due to poor branching performance, like we saw in Flitz Chess

 

[Dresdenboy]

This could tell us that Ryzen's SMT doesn't do as good a job at managing the resources between the two threads.
Makes sense, Intel's years and years of work on SMT basically amounted to minimizing the performance regressions in certain cases.

One problem here is, that those threads do exactly the same thing and need the same resources. This might simply end up in equal sharing.

While the thread confirms that PN score is IDIV based, but should the Integer results not have poor results too? At least this thread seems to suggest it. And if it an error on AMD 12h CPUs, should it now have been fixed by now?

They might have used better code/compilers for PT9. While browsing PM results I saw a significant drop in PN when going from Win32 to Win64.

The error should've been fixed a while back. I think, BD also suffered from that.

But I agree with you on the string sorting and the fight for cache. I also think it might be due to poor branching performance, like we saw in Flitz Chess

Branching should be fine on final silicon. The leaked "fake" CB screenshot might also be real, as the new leak suggests a similar result. AMD might kindly have asked the moderators to call it "fake". Or why would someone delete a photoshopped pic?

 

[Nothingness]

My assembly code was from Intel compiler using MS disassembler. No optimization, as the code is too simple and the optimization will lead to constant as the results donot change with time.

Then what was the aim of your post exactly? That if you ask for a mod operation and don't optimize you'll get a div instruction?

If you want you can use cin to let the user set the values and it should generate the assembly you want even with optimization.

Again the Atkin sieve does a mod by a compile-time know value, so it will be optimized.

EDIT: I read the previous posts that link to passmark forum. How did they achieve to make Atkin sieve limited by modulo operations?!? That is frightening. Talk about useless benchmarks...

 

[Agent-47]

^^ to find out if the compiler uses IDIV or some other fancy instruction? It seems Passmark confirmed it as well

I guess what I am saying is this is to test to check performance of a set of instructions, in this case IDIV, branching and possibly mem. Not the optimization effectiveness of the compilers

 

[SpaceBeer]

Just because AMD didn't have CPU with SMT until now, doesn't mean they are learning now how it works. I'm sure they had many SMT tests, prototypes, simualtions... in last 15 years, or more.

 

[Nothingness]

^^ to find out if the compiler uses IDIV or some other fancy instruction? It seems Passmark confirmed it as well

Yeah I read the rest of the thread after answering you, and edited my post Passmark looks like a very dumb and poorly implemented benchmark...

 

[Dresdenboy]

Here is the DRAM speed chart based on itsmydamnation's data:

This shows, that this factor has a smaller influence than DRAM latency.

I'm a bit late as I got a flu. ^^

 

[OrangeKhrush]

I think the IMC is behind but it seems more like quad channel and no HT is a massive boost.

Performance of the CPU otherwise bis fantastic. Enthusiast hardware at mainstream price

 

[MajinCry]

Any specific workloads (must be available for Windows) you would like to see, besides the current ones?
Preferably open source, but that's not mandatory if the workload can otherwise be justified.

I'd like to see the FPS your CPUs get in NVidia's instancing benchmark, with instancing disabled. See this thread: https://forums.anandtech.com/threads/measuring-cpu-draw-call-performance.2498467/

Download link: http://developer.download.nvidia.com/SDK/9.5/Samples/DEMOS/Direct3D9/HLSL_Instancing.zip

 

[Agent-47]

+1 on draw call benchmark

 

[sm625]

sm625 has gone missing. He unintentionally fuelled the hype train, trolled himself and set us all for a journey to infinity... and beyond

I dont even think I can actually say why I went missing without "going missing" again. My new avatar should explain though. At any rate I'm glad people are looking into that questionable memory score.

There is another passmark baseline now, #774164. It scores roughly the same. I would post some screenshots from passmark, but apparently in north korea, you not even allowed to post screenshots of passmark.

 

[sm625]

i7 5820K @ 4.25GHz - No Turbo
DDR4 @ 3GHz Dual Channel

This is what is scary about Ryzen. Going from 4 channels to 2 channels drops your score from 3300 to 2800. So going to 1 channel would probably drop your score to .... 2200? Still higher than Ryzen. I dont suppose you'd be willing to test that?

 

[Agent-47]

This is what is scary about Ryzen. Going from 4 channels to 2 channels drops your score from 3300 to 2800. So going to 1 channel would probably drop your score to .... 2200? Still higher than Ryzen. I dont suppose you'd be willing to test that?

Quad channel is overrated. You need to stop looking at the total score and the individual scores instead

The scores for individual tests shows only threaded score to be lower. That is a pointless test. During this test a 8c16t CPU will launch 16 threads. All thread allocates a total of 256MB. During the test each thread reads the memory block without doing any processing on it. this is why when you disable HT, this score increases implicating that the higher score can be achieved by lower thread count.

More importantly, how often you planning to run 16 memory intensive tasks that only reads data without any post processing?

 

[Agent-47]

How did they achieve to make Atkin sieve limited by modulo operations?!?

Bit off topic but I just saw your edit, my guess is they made a static class that only takes an externally defined range along which it finds the prime numbers. the compile will not be able to optimize away the IDIV in this case when this class is compiled as it does not know the range. then the "main program" calls the class like a standard static method, i.e. class->calculate_prime(int range).

 

[Face2Face]

I'd like to see the FPS your CPUs get in NVidia's instancing benchmark, with instancing disabled. See this thread: https://forums.anandtech.com/threads/measuring-cpu-draw-call-performance.2498467/

Download link: http://developer.download.nvidia.com/SDK/9.5/Samples/DEMOS/Direct3D9/HLSL_Instancing.zip

Done. I've uploaded it to your thread, but here it is anyway.

 

[KTE]

If dinos is saying this, the performance must be good. I'm a believer

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

I'd like to see the FPS your CPUs get in NVidia's instancing benchmark, with instancing disabled. See this thread: https://forums.anandtech.com/threads/measuring-cpu-draw-call-performance.2498467/

Download link: http://developer.download.nvidia.com/SDK/9.5/Samples/DEMOS/Direct3D9/HLSL_Instancing.zip

Will do, when I start 3D testing.

 

[swilli89]

If dinos is saying this, the performance must be good. I'm a believer

Sent from HTC 10
(Opinions are own)

Who is Dino!? And like the commend underneath him.. I think many of us on 4c/4t Intel builds (2500K) from the last several years will be a prime target for a Ryzen upgrade.

 

[.vodka]

Who is Dino!? And like the commend underneath him.. I think many of us on 4c/4t Intel builds (2500K) from the last several years will be a prime target for a Ryzen upgrade.

An extreme overclocker who's been on the scene for long. Remember many posts from him at xtremesystems ten years ago.

If those guys give Zen their seal of approval... That's it.

 

[Arachnotronic]

Who is Dino!? And like the commend underneath him.. I think many of us on 4c/4t Intel builds (2500K) from the last several years will be a prime target for a Ryzen upgrade.

He's a marketing guy at GIGABYTE.

 

[KTE]

Yea extreme OC since 2005 days on XS, but someone I personally have known... Even in the Phenom/Bulldozer extreme hype days, he was repeatedly attacked for calling them crap, pre-release. Doesn't mince his words.

Not saying its a SKL killer but it can't be bad with his comment.


Sent from HTC 10
(Opinions are own)

 

[swilli89]

He's a marketing guy at GIGABYTE.

Ah ok so that makes sense that he may have some insight into Ryzen I suppose?

Yea extreme OC since 2005 days on XS, but someone I personally have known... Even in the Phenom/Bulldozer extreme hype days, he was repeatedly attacked for calling them crap, pre-release. Doesn't mince his words.

Not saying its a SKL killer but it can't be bad with his comment.


Sent from HTC 10
(Opinions are own)

Certainly seems to bode well!

 

[Arachnotronic]

Ah ok so that makes sense that he may have some insight into Ryzen I suppose?

Of course he does, the major mobo guys have Ryzen.