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

[BeepBeep2]

Noticed in the supposed 6 core picture that it was running at only .374v at 3.4ghz without turbo (unless CPU-Z just isn't reading everything correctly). Color me skeptical on how legit that picture is. Not that I want it to be...just we live in the era of fake news now.

http://pclab.pl/zdjecia/artykuly/blind/2017/02/amdryz/amd582.jpg

Happens quite often, CPU-Z and HWMonitor not reading correct vcore.

That is no basis to call it fake imho.

 

[railroadmaster]

Delete

 

[lopri]

Yes and no. I believe that memory speed/timing combination may have good latency but insufficient bandwidth. One would be better off with DDR4-3200 CAS/CL14 which is still quite nice:

http://pcpartpicker.com/product/jpH48d/gskill-memory-f43200c14d16gvk

Agreed. DDR4-3200/CL14 was the best binned part when I checked out memory sticks, without an absurd price tag. I would also keep an eye on secondary timings as well. tRCD is just as a big part of the equation and a lot of vendors try to get away with just CL ratings only.

@railroadmaster: That is a lovely cobalt blue.

 

[GroundZero7]

@railroadmaster: That is a lovely cobalt blue.

Cobalt blue is quite a bit deeper and darker than that. That is/was more like True Blue.

 

[OrangeKhrush]

I think the Passmark benches were most plausible as the outcome. From a very trustworthy source I'm sticking with my haswell possible slightly faster expectation. Getting to SKL and KBL would require higher clock speed. I would love it to be true but Haswell-Ryzen-Broadwell is far more realistic

 

[inf64]

So close to the launch I think it's time to do a final estimation on Zen's IPC. It is very hard to pinpoint the IPC of a core that has clear compromises in perf. Vs power (ie. slimmer FP unit) but I will give it a try, one final time.

I already tried to do so in this post. Long story short, I expected that overall average IPC is going to be ~46% higher than XV. Since then I have changed my estimate and now I expect Zen to be ~48% faster than XV core.

To put this number in perspective, look at this post from majorD. In his benchmarks Skylake has 1.61x (61%) higher IPC than XV core, both chips tested as dual cores with CMT and SMT disabled, so pure single core IPC comparison. Also I took the numbers that had AVX2 workloads in the average since Zen will be tested the same way. Thanks to Anandtech we know that SB to IB was 5.8% jump, IB to HSWL was 11.2% jump, HSW to BDW was 3.3% jump and BDW to SKL was 3% jump(DDR4). That makes HSW core 1.52x or 52% faster than XV core.

Overall, the gain from SB core to SKL core is around ~25%. Going back to majord's numbers, SB-level of performance would be 1.61 / 1.25 ~=1.29 or SB core has around 29% higher IPC than XV core with AVX workloads in the mix. IB in turn should have around 36% higher IPC and Haswell 1.36x1.112~=1.52x or 52% higher IPC than XV core.

Going back to Zen estimate, ~48% faster than XV would make it a smidge slower than HSW on average (2-3 % points) and around ~6%/~9% slower than BDW-E/SKL, at the same clock. All of this of course without SMT gains taken into account. Looking at CPUz ST->MT scaling on that 6C/12T ES, I can say that there is no surprise in my mind how 3.4Ghz Ryzen can beat 3.5Ghz BDW-E in Handbrake and Blender while having somewhat lower IPC(I guesstimate 6% lower). Average SMT gains on post SB cores are in the 23-25% range. Ryzen could very well have ~35% gain from its SMT implementation making up for the pure ST IPC deficit vs BDW and SKL cores. 1.35/1.25=1.08 or 8% more performance from their own version of SMT,and like I said before, just about enough in couple of workloads we have seen thus far (and reinforced by higher-than-intel CPUz MT scaling on R5 12T ES).

This was fun for me but, as always, take it with healthy amount of salt of course

 

[ultima_trev]

Between Canard PC, Passmark CPU Mark and 3DMark Firestrike Physics leaks it seems safe to say the $400 R7 1700X will offer 90% of the i7 6900K's performance overall. Zen seems to be very competent in Compression, Encryption, Integer Math, Floating Point while single thread is just a few percent behind Broadwell-E. Prime Numbers and Physics probably won't affect the overall too much as it's inline with Intel's high end quad cores.

Tackling Intel's HEDT offerings seems like it should be easy.

However the Kaby Lake 7600K and 7700K can easily overclock to 4.8+ GHz and unless Ryzen can do the same gamers will likely favor those while power users who do more than just game will favor Ryzen.

 

[OrangeKhrush]

I will stick with my long standing position since the New Horizons launch. Haswell IPC baseline, with it slightly higher given it can be binned higher and at best mid way to Broadwell.

 

[inf64]

Remember this picture?

From here you can find a link to a screen grab of a cooler used to achieve whatever these 3 were looking at (supposedly):
"The guy on the left shown the cooler used in the test."

All allegedly of course. That cooler doesn't seem like it could dissipate much heat IMO. But who knows.

 

[blublub]

Remember this picture?

From here you can find a link to a screen grab of a cooler used to achieve whatever these 3 were looking at (supposedly):
"The guy on the left shown the cooler used in the test."

All allegedly of course. That cooler doesn't seem like it could dissipate much heat IMO. But who knows.

What's it with that picture? It could just be made to from a couple of guys just keep us busy here

 

[inf64]

What's it with that picture? It could just be made to from a couple of guys just keep us busy here

They were looking at Ryzen Cinebench score @ default as per Alva guy:

Guy in the middle is OCing champion as per users on reddit.

 

[OrangeKhrush]

I am going with Haswell IPC but with some advantages.

- smaller node
- insane TPD
- 6 and 8 core on mainstream platform
- 40+% generational leap
- Prices that are almost unbelievable (I bought my 5960X for 900USD 2 months ago)

After the Kanter interview I think AMD put a lot of effort into this and it will be rewarded.

 

[PhonakV30]

I am going with Haswell IPC but with some advantages.

- smaller node
- insane TPD
- 6 and 8 core on mainstream platform
- 40+% generational leap
- Prices that are almost unbelievable (I bought my 5960X for 900USD 2 months ago)

After the Kanter interview I think AMD put a lot of effort into this and it will be rewarded.

wow crazy expensive ! it's unbelievable that AMD will sell CPU Ryzen to costumers for 1/3 Price with same performance.

 

[Dresdenboy]

If this kind of (or similar) ST performance is reflected in most desktop workloads, I will dump my Haswell 4690K in a blink of an eye for 12T Ryzen. It will cost me some 500 euros but I can maybe get 350 for my current setup, so 150e for 50% more cores and similar ST perf. with smallish OC is a great deal.

Right now you could get 350€ maybe. But what will you probably get in a few weeks?

 

[Dresdenboy]

So Skylake IPC already with CPU-Z bench. Still time to pass it before release. Choo choo!

Then we can all wonder why we keep being shown small cache loop benches.

The CPU-Z benchmark was a hot topic here before. It might only reveal a rather small part of the whole picture.

 

[Dresdenboy]

Well, considering that presumably Dolphin benefits from FMA, that could the explanation.

Do you know what it would use these instructions for?

 

[William Gaatjes]

Single digit slower than the latest Intel design without xfr. That has always been my prognosis.

 

[bjt2]

Pretty impressive it can maintain stability at such low voltage at that clock speed. The norm is that voltage scales down as clock speed scales down. I've seen clocks scale down and voltage stay up...but never voltage scale down and clocks stay up.

According to Bristol Ridge paper, the minimum Vcore for cache retaining on the 28nm BULK is 0.55V. Maybe on 14nm FinFet is lower... And on Zen they modified the cache architecture to have even lower Vcore at same process, so adding up these two informations, 0.375V can be possible, even if I think it is a misreading because on Bristol Ridge they applied a guardband, keeping Vcore no lower than 0.7V, so that with Vdroop it does never go under 0.55V...

 

[bjt2]

Remember this picture?

From here you can find a link to a screen grab of a cooler used to achieve whatever these 3 were looking at (supposedly):
"The guy on the left shown the cooler used in the test."

All allegedly of course. That cooler doesn't seem like it could dissipate much heat IMO. But who knows.

According to a facebook post, the reaction was to cinebench score at DEFAULT (i assume with XFR enabled and that cooler)

EDIT: i arrived too late...

 

[OrangeKhrush]

Still going with a good source either side of 5% to Broadwell

 

[bjt2]

Wait. . .



Ahem.



*cough*



Sure beats the alternative.

The acronym stands for Digital LDO, that can also be written as DiLDO...

 

[Greyguy1948]

The CPU-Z benchmark was a hot topic here before. It might only reveal a rather small part of the whole picture.

dresdenboy.blogspot.com

Yes it doesn't stress L3 cache in a realistic way.
Compare these:
Sandy Bridge i3: ST=1245, MT=2547
Athlon 845 (Exc) ST=1187, MT=4549
Compare with game performance!

 

[Lepton87]

Yes and no. I believe that memory speed/timing combination may have good latency but insufficient bandwidth. One would be better off with DDR4-3200 CAS/CL14 which is still quite nice:

It is quite tight. There was this popular formula effective frequency/cas latency. Anything over 200 was considered tight

 

[CentroX]

Crazy turnaround for AMD

http://www.alphr.com/features/372859/amd-what-went-wrong

 

[Dresdenboy]

Yes it doesn't stress L3 cache in a realistic way.
Compare these:
Sandy Bridge i3: ST=1245, MT=2547
Athlon 845 (Exc) ST=1187, MT=4549
Compare with game performance!

The CPU-Z benchmark makes heavy use of IDIV. And guess what? XV with 4 int cores has two times as many IDIV units.