Discussion: AMD Ryzen

Page 21 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.


Concur; I was hoping the gap would be "close enough", but I agree the multithreaded results should be discarded. So looking at the Single-Thread results:

Zen: http://browser.primatelabs.com/v4/cpu/105227
Ivy: http://browser.primatelabs.com/v4/cpu/42743
Haswell: https://browser.geekbench.com/v4/cpu/233841
Haswell: https://browser.primatelabs.com/v4/cpu/117877

So, one FINAL time, Single Core IPC Comparisons dis-including the AES and Memory benchmarks:

Zen IPC Vs Ivy Bridge [Xeon E7-8857 @ 3.60 GHz]: 22.52%
Zen IPC Vs Haswell [i7-5820K @ 1.40 GHz]: 44.13%*
Zen IPC Vs Haswell [Xeon E5-2603 v3 & 1.60 GHz]: 48.22%

*Geekbench reports 3.30 GHz, but Scuzzycard claims he clocked it at 1.40 GHz. The numbers appear to back Scuzzycard, as the CPU would be underperfoming badly if truly clocked at 3.30.

The fact the HW scores are close shows the math is correct; I'd expect the Xeon to do slightly better then an i7, and it shows here. These are as legitimate as we can get, since the IPC equation reduces to a simple to compute IPC = Perf / Clock. Then you simply solve for % Difference between the average IPC for each CPU.

GeekBench may be overly sensitive to things such as CPU cache. Still, this does show what most of us were expecting: Zen comes *close* in IPC to Ivy Bridge, but doesn't surpass it. I'll test my 2600k this weekend if I have time, but I suspect Zen is slightly lower then SB in single core IPC.

So long story short: Unless Zen is clocked higher then some of us expect, it will likely be slower then a stock Sandy Bridge i7. Again, assuming these numbers are legitimately from a Zen CPU.
 
I noticed another interesting bit of information: "Ubuntu 16.04 LTS 4.4.0-21-generic x86_64".

That is Ubuntu using the generic Kernel. From what I noticed all these years of using Linux, the generic kernel takes some performance away. I can't put a number to it, but maybe it's not that much nowadays... In any case, I don't think that Ubuntu has Zen patches in it, since it's an ES, lol.

Another bit is this: "2S1451A4VIHE4_29/14_N".

2.9Ghz turbo and 1.4Ghz base? What does the "N" stand for?

Cheers!
 
Right, it's quite possible there's some loss in the OS, but to be fair, single-core performance is much less likely to be affected significantly then multithreaded. The numbers are likely close to what the final results will look like, at least as far as Geekbench goes.
 
Yeah, I think you are correct in that regard, gamerk. It was more in the lines of nitpicking information than bluntly saying the results are skewed by it.

The only thing I have in mind is all the little differences we can spot, might be adding into the performance gap you guys are discovering. In a future take, doing this in retrospective, we might be able to extrapolate more information out of the same bits and pieces from half-done benchmarking in leaks. That is, when we get Zen and have a proper go at it, we can go back to this and see how far off the numbers where and bag them in all this little nitpicks.

Cheers!
 
Agreed. It's one thing to look at numbers, but it's another to understand them. Example, the Multithreading numbers are so bad, it's clear something besides the HW is playing a part. So, we disregard them. That's how it works. But if you don't make an attempt to understand the results, you accomplish nothing.

What I CAN say is, at least in Geekbenches test suite, Zen is slower then Ivy Bridge core for core, but likely comes close to Sandy Bridge. That means, as far as gaming is concerned, AMD has reached parity with Intel. For other processes, performance will depend entirely on scaling, as with "only" a 20% IPC deficit, one would mathematically expect AMD to start beating i7 processors in tasks that scale beyond 10 or so cores.
 
To be honest, I don't think "gaming" is one of the things Zen will "shine" in. Given the speed clocks information we have so far, it won't make a dent in the mainstream market. Well, with the 8C/16T part clocked under 3.6Ghz.

Overclocking potential is absolutely unknown and power figures implied here and there for the parts don't paint a really good picture for it. I would imagine 4C/8T will fare much better, since Intel currently has the 4C/8T at 4Ghz around 85W TDP. So AMD has some leeway there, but I don't know how much from the 65W AMD-TDP. I wish they could announce the speeds soon. I think we'll be able to connect more dots by then.

And just to clarify; this train of thought is for the mainstream market (gaming?). I do believe Zen will do fine in "prosumer".

Cheers!

EDIT: Typos.
 


Naples
 


Not so. We know the 14LPP process and Polaris. Overclocking capacity for Zen will range from mediocre to bad.



My guesses

8C/16T 95W: 3.0GHz base 3.4GHz turbo
4C/8T 65W: 3.3GHz base 3.7GHz turbo
 


My thought is this: If IPC is within about 20% of IB, that implies it's within 10% of SB. And at ~3.2 GHz clocks, Sandy Bridge isn't CPU bottlenecked. Therefore, Zen clocked at 3.2 GHz+ will almost certainly not be a bottleneck in the same way Steamrolller is, and will likely produce the same or ever so slightly lower FPS then Sandy Bridge and later CPUs. Point being, in gaming, AMD likely closed the gap.
 
Ok So I've done some back of a napkin math and the Zen scores look too low to me- looking at single core score for the old FX 8370- which gets a single core score of 2650 @ 4ghz, then at 1.4ghz it should get 927.5, however as we've established scores at low clocks should be about 20% faster proportionally, which makes it 1113. That would put the single core IPC of Zen and Piledriver about the same. Given the hardware improvements shown (including the lack of CMT, the MUCH wider execution engine, uOP cache and other enhancements) that doesn't make sense.

Obviously this is based on pretty rough math but I think it's fairly safe to assume that Zen, clock for clock, is certainly faster than PD.

To back this up, when I get home later I'm happy to run Geekbench on my FX 8320 clocked at as close to 1.4ghz as I can get it to give us a comparison point. If it scores close to Zen then the Zen results are either wrong (due to a technical issue with the platform) or outright fake.
 


Here you can find score for FX-8370 @4.02GHz

http://browser.primatelabs.com/v4/cpu/230545

The score is lower than yours due to margin of error. Moreover, don't forget turbo effects as well.

About Zen being always faster, well Steamroller microarchitecture was also very improved over Piledriver, still I can show you a pair of benchmarks where Steamroller was slower than Piledriver clock-for-clock

Richland-Kaveri-Comparison.png
 


Uhm... It would be weird if GF tuned the process the same way for a CPU and a GPU. Think about it this way: if it was tuned for CPU speed, then why Polaris has such low clocks?

And this reasoning would also imply AMD *chose* to use a process type that won't clock high for a CPU that needs as much speed as it can get (reasonably).

I would like to extrapolate from what we have seen of Polaris, but it will really go down to how GF has tuned the process for the CPU.



Uhm... Half the CPU resources disabled, 300Mhz extra (same voltage?) and 30W less... I really don't know how to blend them into a projection without knowing the line voltage for the 95W part. Has anyone seen the voltages?



That is a good point. There's a reason we have recommended i3s for a while now as viable gaming CPUs. Maybe AMD won't be in such a bad standing... I really hope so at least.



There's also the thing about what instruction set Geekbench is using/testing for it's main threads. If it's using, say, SSE3, then I really don't know how to compare them there. Would PD have the same SSE3 performance as Zen? I would imagine no, since it was way better cache and a bloody uOP cache now. Would AVX perform the same? Same rough answer.

But please do test, cdrkf. More information never hurts :)

Cheers!
 


Interesting numbers. I imagine we shall see sooner rather than later what it truly is.

As for the clock speed, most benchmarks I have seen read the processor spec info. Most say my 4670K is at 3.3GHz but I have it at 4GHz right now.
 
Ok so I've run geekbench 4 on my Fx 8320. Locked the multiplier to 7 for 1.4ghz and disabled turbo core. I monitored the test and CPU speed stayed at 1.4ghz the whole time, although the tool reports a clock speed of 2.9 ghz (I tested at normal as a control and the scores suggest it was, in fact, at 1.4ghz however). Link to the result:

https://browser.geekbench.com/v4/cpu/243008

So, Piledriver appears to score 1024 at 1.4ghz in the single thread test. That is awfully close to Zen, by my reconing it would put Zen at 12% more IPC than *Piledriver*- I'm fairly certian Steamroller and certianly Excavator already have that kind of an advantage. I think these results *must* be fake- or at the very least this is a really early ES part that has some features disabled. AMD are claiming 40% more IPC than Excavator- Zen should easily score 40% higher than a PD part and really should be in the 50 - 60% faster region if the 40% boost over EV is true.

out of interest- if we say conservatively that a fully working Zen part is 40% faster than my CPU (which would be under AMD's official target), in this test we should see a score of 1433- which is faster than IVY and close to Haswel.

TL, DR; Zen benchies are blatantly wrong, likely fake or possibly on a restricted early ES part. Nothing to see here 😛

Edit: Control test at full speed:
https://browser.geekbench.com/v4/cpu/243304

Also I notice my memory speed was set lower than it should be for the 1.4ghz test- so that result is slightly lower than it should be which brings this cpu even closer to the 'zen' sample.
 
Has anyone come to the conclusion that its probably fake to begin with lol, i said it before its either fake or something is wrong no way it would score that even a piledriver CPU with that many cores would fair better.
 
Uhm...

Zen leak:
Memory Latency -> 1381 - 3.19 Moperations/sec

8320:
Memory Latency -> 3759 - 8.68 Moperations/sec

There's something fishy there with the memory settings. The leak shows a 128GB of memory and assuming it's a server class testing board, would they be testing a dual quad channel board with a single 128GB stick? or maybe 2x64GB in single channel config per CPU? I know that will skew some of the results badly, but other than the explicit memory tests, can't say for sure which others will.

Interesting information indeed.

EDIT: Thanks for the info, cdrkf!

Cheers!
 


The only 128GB DIMMs are ECC which would make it slower than regular RAM. I can't phathom why they would do a test in single channel mode on slower RAM.
 


Nope.

Haswell Xeon @1.6GHz:
Memory Latency -> 2324 - 5.37 Moperations/sec

This subtest is very sensible to clocks and prefetching effects.

What is really catching my attention is how Geekbench reports the L3 cache

"L3 Cache 65536 KB x 8"

This suggests that the L3 slices within each Zen complex are independent. :ouch:
 


It is weird how it reports the L3, yes.

And the 8320 was also clocked at 1.4Ghz. That is why it caught my eye. The IMC design from DDR3 to DDR4 can't change that much, can it?

Cheers!
 


If the L3 is a unified LLC, then I agree that GKB reporting is weird, but if the L3 is not unified then GKB is doing it fine.

DDR4 has been designed for higher throughput at cost of latency. Latency has increased a lot of until the point that DDR4@2400MHz is slower than DDR3@1333MHz

Memory%20True%20Latency.png
 

So it's an inverse scale? Well, that would make sense then.

Cheers!
 
http://wccftech.com/amd-zen-intel-kaby-lake-february-2017/

The gist of this post is that Zen desktop is coming in Feb 2017. Also, some Kaby Lake news. We should have a lot of answers by the end of February.
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom