ZEN ES Benchmark from french hardware Magazine

Page 25 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
I went back to CPC's claim of 35% higher performance of Zen 1C/2T Vs PD 1C/2T. It turns out that according to this we have A0 sample they had tested achieves around 1.69x IPC without SMT Vs PD core(without its CMT). Hardware.fr found out in their initial FX8150 review that average gain from CMT is around 1.53x, while SMT on intel SNB core netted 23.4% on average. I assumed that Zen will be no worse than SNB when it comes to SMT gains. Pilderive core is baseline 1.
One core PD performance without CMT: 1/1.53=0.65
One core Zen performance without SMT: 1.35/1.23~=1.1 => Zen has 1.1/0.65~=1.69x or 69% higher IPC.

From AT Carrizo review we have generational performance uplifts calculated for PD, Kaveri and Richland. In summary from PD core to Carrizo core we have around 15.5% more IPC (average figure from apps and games). Now we can see how much better than Carrizo is the A0 sample that CPC tested: 1.692/1.155~=1.46 or some 4.6% better than what AMD claimed. If there are any performance hindering bugs on that sample I don't expect them to affect the IPC by more than 5% (from A0 to whatever retail stepping we end up with). In conclusion, I think at best the IPC will be 53% better, at worst we are having around 45% better IPC - the average figure of course. Maximums and minimums will likely exist and will vary a lot from benchmark to benchmark. Overall the core seems to be very strong and close to Haswell/Broadwell level. I expect 4-5% overall lower than Broadwell-E and around 6-7% lower than Skylake level, basically AMD's Haswell version.
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
I went back to CPC's claim of 35% higher performance of Zen 1C/2T Vs PD 1C/2T. It turns out that according to this we have A0 sample they had tested achieves around 1.69x IPC without SMT Vs PD core(without its CMT). Hardware.fr found out in their initial FX8150 review that average gain from CMT is around 1.53x, while SMT on intel SNB core netted 23.4% on average. I assumed that Zen will be no worse than SNB when it comes to SMT gains. Pilderive core is baseline 1.
One core PD performance without CMT: 1/1.53=0.65
One core Zen performance without SMT: 1.35/1.23~=1.1 => Zen has 1.1/0.65~=1.69x or 69% higher IPC.

From AT Carrizo review we have generational performance uplifts calculated for PD, Kaveri and Richland. In summary from PD core to Carrizo core we have around 15.5% more IPC (average figure from apps and games). Now we can see how much better than Carrizo is the A0 sample that CPC tested: 1.692/1.155~=1.46 or some 4.6% better than what AMD claimed. If there are any performance hindering bugs on that sample I don't expect them to affect the IPC by more than 5% (from A0 to whatever retail stepping we end up with). In conclusion, I think at best the IPC will be 53% better, at worst we are having around 45% better IPC - the average figure of course. Maximums and minimums will likely exist and will vary a lot from benchmark to benchmark. Overall the core seems to be very strong and close to Haswell/Broadwell level. I expect 4-5% overall lower than Broadwell-E and around 6-7% lower than Skylake level, basically AMD's Haswell version.

On PD the average CMT yield is >= 75%.
What A0 stepping Carrizo are you talking about? AT used Athlon X4 845 retail part for the review, which features the only available final silicon version (CZ-A1).
 
May 11, 2008
19,303
1,130
126
Well, I'm walking into somewhat of a minefield here, but the AMD/GCN cards do have a small design "fault". It's not relevant for gaming, but they are quite power hungry during video decoding, because AMD chooses to run the memory at full speed. Again this is not an issue as such if you use your PC for gaming, but there are a few situations where it is relevant.

power_bluray.png


I hope AMD/RTG will be able to finetune that with a driver update.
I am going to disable my igpu when i receive the rx480 card. Devastator igpu is pre-gcn and not compatible with rx480.
 

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
On PD the average CMT yield is >= 75%.
What A0 stepping Carrizo are you talking about? AT used Athlon X4 845 retail part for the review, which features the only available final silicon version (CZ-A1).
Maybe I phrased it wrong, A0 was referring to Zen part used by CPC in their review.

As for CMT yield I used hardware.fr results from their Bulldozer review. I doubt that PD core has any better CMT performance, AMD just upped the IPC of the core by around 7-10% from what I saw in the reviews.
Edit: Are you sure that average CMT yield is the 75%. Hardware.fr found that many MT workloads didn't see more than 50% (or even less than that):
bNqCsUp.jpg


Eliminating all the games(ST) and low score outliers ( like Winrar/1st pass) I get 49% average CMT yield.
 
Last edited:

Abwx

Lifer
Apr 2, 2011
10,853
3,298
136
Hardware.fr found out in their initial FX8150 review that average gain from CMT is around 1.53x,

Those numbers are outdated, PD gain in FP can be quite big and CPC used mainly FP benches, FTR they used Mental Ray within 3DS MAX wich is also used by HFR, in this test PD has 10% better throughput/Hz than BD.

http://www.hardware.fr/articles/880-6/bulldozer-vs-piledriver-4-ghz.html

Those numbers include the CMT scaling, and The Stilt estimation is closer to the actual numbers, i would say 70% in FP...
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
The most recent info about the CMT yield on PD I found is the ones posted by superstition, where he tested it using the different Blender builds. For the properly optimized builds it is around 68%. These can be found in the recent pages of the original Summit Ridge thread.
 
  • Like
Reactions: inf64

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
Well if I take the 70% average CMT gain we would end up at 1.88x IPC gain over PD and 63% gain over XV core which I believe is not realistic. Unless of course Zen has 35% gain from SMT(not impossible) in which case we are back at similar numbers as in my original post: ~72% over PD, 49% over XV. So either SMT gain on Zen is a lot greater or CMT gain on PD core is not ~70% on average. Former is likely to be correct as there are speculations that Zen sees more SMT gains due to FP co-processor model AMD used.

Edit:

I went back to games test in CPC preview. 8C Zen Vc 8C PD, so any small effect that MT games have on the average score is already calculated in and both chips have good MT support (so it's not like comparing i5 to i7). I averaged the clock for both and didn't take the best clock they could achieve due to nature of some games and early stage of the whole Zen platform.
3.3Ghz Zen -97.3
4.1 Ghz PD - 73.6 => Assuming linear increase in performance (not likely) Zen at 4.1Ghz could score 120 which is 64% higher.

And before someone here jumps on me on how the games won't scale with clock, just compare 4790K below vs 4770K :

tU8O8jj.jpg


4790K Haswell 4C/8T 4-4.4Ghz range, average 4.2Ghz - 136.4 pts
4770K Haswell 4C/8T 3.5-3.9Ghz range, average 3.7Ghz - 121.7 pts
Difference in performance is 136.4/121.7=12%. Difference in average clock is 13.5%. Almost 100% scaling, to be precise 98.7% ;)
 
Last edited:

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
Well if I take the 70% average CMT gain we would end up at 1.88x IPC gain over PD and 63% gain over XV core which I believe is not realistic. Unless of course Zen has 35% gain from SMT(not impossible) in which case we are back at similar numbers as in my original post: ~72% over PD, 49% over XV. So either SMT gain on Zen is a lot greater or CMT gain on PD core is not ~70% on average. Former is likely to be correct as there are speculations that Zen sees more SMT gains due to FP co-processor model AMD used.

Do the figures you mention refer the workloads in the chart you posted?
 

inf64

Diamond Member
Mar 11, 2011
3,685
3,957
136
Do the figures you mention refer the workloads in the chart you posted?
Yes, for BD they do. Also I assume that those workloads are similar to what we have seen in Zen preview. The workloads are not the same of course (hardware.fr Vs CPC's preview) but the performance picture will not vary that much.
 

majord

Senior member
Jul 26, 2015
433
523
136
Perhaps I've missed something, but there doesn't seem to be any indication that Ryzen actually
matches BW-E IPC on avg. Rather is slightly behind.. Enough to be considered more like Haswell in throughput, or slightly lower.. This is based on the Canard tests which show a 6900K at a 14.5% advantage, (including blender) and which, by all accounts has a <10% clock advantage. So perhaps people shouldn't be getting there knickers in a twist 're who's predictions were right or wrong quite yet? Moreover, the 40% is an ST claim, and none of the benchmarks so far have been run 1T per core. So until someone does this, no one knows.

By my own testing of Excavator in 1T/module vs Skylake i3 1T/core , I stiill avg around 60% to SKL no matter what I add or remove from the bench set to try and get a realistic avg (which is difficult due to the wildly different architectures), that puts Skylake approx 15% > A theortical Zen. Which in turn plants it pretty firmly in between ivy and Haswell.

A bit more than The Stilt's original prediction, and still sounds low to some, but personally I still predict Ryzen will have greater SMT Yeild (%) than Broadwell, which means of course that throughput can be closer to Haswell/BW than it's ST performance is.

As for why I think this will be the case:- Distributed Int Schedulers, and a higher number of statically partitioned structures are the main reasons. Intel's more dynamic approach here should result in higher ST performance, but IMO is a mistake in regards to throughput perf/watt. if not perf/mm , It makes sense to me that AMD chose to avoid such a large unified scheduler.. This is after all an architecture designed to be as balanced as possible, But given it's 2017 now (Is here in Australia!) likely biased towards throughput wherever it didn't impact ST significantly,


rZ734LS.png
 
Last edited:

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
Yes, for BD they do. Also I assume that those workloads are similar to what we have seen in Zen preview. The workloads are not the same of course (hardware.fr Vs CPC's preview) but the performance picture will not vary that much.

Ok.
We'll see ;)
 

KTE

Senior member
May 26, 2016
478
130
76
Flanker: have you won a Ryzen giveaway?!?

I hope this won't cause another class action suit. ;) And what should we tell the "up to 40%" believers?.
If AMD beats thats IPC, without SMT, like they said, then all that can be said is kudos, that they've made a revolutionary and impressive design. K8 v2.

Heck, even >30% IPC on average would be amazing!

But the question still remains unanswered: if that really is IPC over EXC they mentioned, or, throughput

The other question remains of average IPC... I have always expected lower than Broadwell IPC but higher SMT efficiency.

The bug (often compared to the TLB bug) might cause SB level IPC then. Any leaks out there might also suffer from this.
Good to know, but again, this is something AMD should control and remedy before launch. Not for an end user or reviewer to play around with.

It looks like it has 7 chokes / phases at first, but there is an eighth choke in there.

It's kind of weird. This looks like a mid range board at best, nothing too special
Quality components and I'd have no issues with even 4+2 designs. Even boards like the Asrock 890GX (4+1), we ran 1.5v 9850BEs 24/7 std aircooling for nearly a year no problem. That was 90+20A stock, so you were looking at >200W after OC there.

Do we have any information on what voltages or power planes Zen products need to work? This being a socket that can support APUs or CPUs, both being SoCs with a southbridge/FCH in there too..
In particular, what are their Idd limits for each core/component. Low Idd and such VRMs would be fine but I doubt Idd will be low.

Idd scales with voltage.

That appears to be a pretty standard mainstream board, not the high-end (UP-series) we've seen before.
4+3 was chosen as one-size-fits-all type of solution, which isn't really good for anything else but the 65W Bristol and Raven Ridge APUs.

We definitely do not want 8+2 phase boards, unless we have no intention of using the same board with Raven ;)
So the GPU side must be pretty darn powerful to require at least 3 MOS. Good to know!

Sent from HTC 10
(Opinions are own)
 
Aug 11, 2008
10,451
642
126
what's the problem with Polaris? It is already better than its more expensive competitor, the 1060 6GB. Yes, there was TONS of hype before Polaris, but look at where it is now, and what people were saying about "where it would be" only 6 months ago.

I'm actually impressed.
Despite the concerted effort in these forum to discredit it, power consumption does matter, and Polaris was a dud in this area. It was especially disappointing after all the hype by AMD (and the fans in this forum) about 2.5x performance per watt improvement, how it would beat nVidia in efficiency, blah, blah, blah, when the majority of cards failed to even come close to that. In fact, despite AMD's backtracking that it was for only one model, I dont know if any model ever reached that efficiency. It was also quite entertaining to see all the posts about how wonderful polaris would be in performance per watt, and the desperate backtracking and reversals of field when it turn out nVidia was still more efficient and we were then told to change our light bulbs and quit drinking beer instead of buying an efficient card.
 
  • Like
Reactions: Sweepr

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
So the GPU side must be pretty darn powerful to require at least 3 MOS. Good to know!

Yes and no.
Things are much more complicated with Zen based products, as the requirements can change quite heavily between different settings...

Of course it is possible to support Raven with 2-phase secondary plane, however it would require the use of high-end components which are significantly more expensive than three or even higher number mainstream quality phases.
Most of the mainstream boards will most likely utilize the new controller family from Intersil, which supports flexible phase configuration. They're the only controllers which support more than two native secondary plane phases, so they are an convenient choice for AM4 boards.
These controllers support up to 7 phases, which means that the most common configuration will probably be the same 4+3 seen on the Gigabyte. Higher-end boards most likely go with a doubled digital configuration.
 
  • Like
Reactions: KTE

Abwx

Lifer
Apr 2, 2011
10,853
3,298
136
But the question still remains unanswered: if that really is IPC over EXC they mentioned, or, throughput

That s about the 100th time that such a "question" was answered in this very thread, do you know that we have the info since Hotchips..?.

It was especially disappointing after all the hype by AMD (and the fans in this forum) about 2.5x performance per watt improvement, how it would beat nVidia in efficiency, blah, blah, blah,

You should open a thread in the VCG section, this would spare us yet another thread derailing post of yours, the thread is about Zen and there s no GPU in the story, much less Nvidia CPUs to compare...

Edit : on second thoughts and considering that your post came long after the initial derailing i m inclined to think that it is a willfull thread crapping/derailing, i guess that we had to endure this sabotage till the last day of the year, so much about talking of rabid AMD fans, at least you wont see them constantly filling Intel related threads with whatever is irrelevant.
 
Last edited:

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
But the question still remains unanswered: if that really is IPC over EXC they mentioned, or, throughput
We've been over this, if it were +40% throughput then IPC gain would be even higher.
 

bjt2

Senior member
Sep 11, 2016
784
180
86
Zen gain in SMT could and should be higher than INTEL's because Zen can do 4 int + 4 fp + 2 mem, while intel can do 4 int OR 3 FP + 4 mem... Two fat threads will fight for the int and FP ports on INTEL, while running smoothly on Zen...
 

Thedarkchild

Junior Member
Dec 30, 2016
7
16
16
Despite the concerted effort in these forum to discredit it, power consumption does matter, and Polaris was a dud in this area. It was especially disappointing after all the hype by AMD (and the fans in this forum) about 2.5x performance per watt improvement, how it would beat nVidia in efficiency, blah, blah, blah, when the majority of cards failed to even come close to that. In fact, despite AMD's backtracking that it was for only one model, I dont know if any model ever reached that efficiency. It was also quite entertaining to see all the posts about how wonderful polaris would be in performance per watt, and the desperate backtracking and reversals of field when it turn out nVidia was still more efficient and we were then told to change our light bulbs and quit drinking beer instead of buying an efficient card.
Its still faster chip. With more memory, more TFLOPs and cheaper price. It uses 40W more? Well that sucks, but it sucks more if AMD wants to deliver the goods in laptop. If I'm building budget gaming PC I'll take faster card that is more future proof.

BTW, the biggest issue with perf per watt for Polaris was that GloFo simply didn't deliver. XFX version of the chip is latest AIB and is comfortably best when looking at perf per watt, plus, its overclocked out of the gate.
 
May 11, 2008
19,303
1,130
126
I will stick to the thought where Zen will be in the single digit percentage slower on average than its designated Intel adversary.
And that will be good enough.
 

KTE

Senior member
May 26, 2016
478
130
76
As for CMT yield I used hardware.fr results from their Bulldozer review. I doubt that PD core has any better CMT performance, AMD just upped the IPC of the core by around 7-10% from what I saw in the reviews.
Edit: Are you sure that average CMT yield is the 75%. Hardware.fr found that many MT workloads didn't see more than 50% (or even less than that):
bNqCsUp.jpg


Eliminating all the games(ST) and low score outliers ( like Winrar/1st pass) I get 49% average CMT yield.

Keep in mind that they are measuring high Core/Thread scaling rather than CMT gain there. Scaling to 8 thread from 4 is typically poor anyway, so that test limitation would give us a pretty awkward result.

I'd say trying 2->4 would be a better indicator.

Keep it mind its entirely possible for Zen to perform +100% to Piledriver in some benches. There's always uarch favoring loads, and 128b sp fp seem like a good candidate

Sent from HTC 10
(Opinions are own)
 
Last edited:

KTE

Senior member
May 26, 2016
478
130
76
That s about the 100th time that such a "question" was answered in this very thread, do you know that we have the info since Hotchips..?.

Yes, you're right. They answered the question about their aims -> per core. I totally forgot.

Sent from HTC 10
(Opinions are own)
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Of course it is possible to support Raven with 2-phase secondary plane,

Okay, I'm curious: how? Do you have to split one plane between the NB and GPU?

Most of the mainstream boards will most likely utilize the new controller family from Intersil, which supports flexible phase configuration.

Interesting. So those 4+3 boards can commit one extra phase to the CPU for Summit Ridge . . . wait, what's with the extra choke on that Gigabyte board, though?
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
Okay, I'm curious: how? Do you have to split one plane between the NB and GPU?



Interesting. So those 4+3 boards can commit one extra phase to the CPU for Summit Ridge . . . wait, what's with the extra choke on that Gigabyte board, though?

CPU cores are fed from one plane and all of the other high power parts from the SoC plane. Carrizo / Bristol Ridge requires three planes by the design, however on AM4 BR parts both of the SoC & GFX planes are fed from the same plane as AM4 has only two major power planes. This required some software configuration changes to BR and hurts the overall efficiency slightly. Things are not as simple with Zen based parts ;)

The phase configuration obviously cannot be changed after the design has been done. The specific controller just allows more flexible configuration (i.e same controller can be used with e.g 6+1 and 1+6 phase boards). The Gigabyte board uses 4+3 configuration.