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

[USER8000]

You're really not going to get anything legit until reviews.

Well,I thought with the fact it did not say "rumours" it was some more benchmarks that had been released - in the past tech sites have done Intel and AMD sanctioned previews before launch.

Edit to post.

Also,thinking about the leak - if this was one of the very low clocked validation CPUs,running at under 3GHZ,then that Intel chip will have at least a 15% to 20% clockspeed advantage.

 

[BeepBeep2]

Well,I thought with the fact it did not say "rumours" it was some more benchmarks that had been released - in the past tech sites have done Intel and AMD sanctioned previews before launch.

Edit to post.

Also,thinking about the leak - if this was one of the very low clocked validation CPUs,running at under 3GHZ,then that Intel chip will have at least a 15% to 20% clockspeed advantage.

Unless it was 2.1/2.4 GHz base and 3.5 GHz boost it was most likely at 3.1 or 3.4 GHz base clock.
I mean, who knows what the actual clock speed was at the time the benchmark was run, but there is enough showing there to determine the most likely clock speed of the ES was X.1 / X.4 GHz base and X.5 Ghz boost.

 

[USER8000]

Unless it was 2.1/2.4 GHz base and 3.5 GHz boost it was most likely at 3.1 or 3.4 GHz base clock.
I mean, who knows what the actual clock speed was at the time the benchmark was run, but there is enough showing there to determine the most likely clock speed of the ES was X.1 / X.4 GHz base and X.5 Ghz boost.

The validation samples had a 2.8GHZ baseclock according to the previous leaks - AMD said the Ryzen sample they used had no functional Turboboost and they set the voltage manually for it to run at 3.4GHZ!

If samples are going to leak,it is going to be those and if AMD can't get Turboboost running in time for the demo,I doubt some Chinese bloke on a forum could.

Edit to post!

Here is one of the many articles mentioning it:

http://www.kitguru.net/components/cpu/matthew-wilson/amd-zen-engineering-sample-benchmark-leaks/

Now,consider if that Ryzen sample was running at under 3GHZ,the Intel would be running at 15% to 20% higher clockspeed with quad channel RAM.

 

[USER8000]

Also,we are assuming AMD SMT GEN1 is going to scale as well as Intel SMT GEN1000. Intel have had many more years of using SMT than AMD has.

Edit to post.

Speculating for a bit - if the validation CPU was stuck at 2.8GHZ,instead of 3.2GHZ,the Intel Core i7 6900K would be scoring 30% more for 25% greater clockspeed,assuming Turboboost is halfway between the base and boost clockspeeds for the Intel chip.

 

[Doom2pro]

Middle is original image stretched vertically, bottom is typed in dresdenboy guessed number, above white on black is a portion of the top of that stretched similarly to original.

The font used wasn't exactly the same (not sure what CB uses for font)... Take from it what you like.

 

[USER8000]

Middle is original image stretched vertically, bottom is typed in dresdenboy guessed number, above white on black is a portion of the top of that stretched similarly to original.

The font used wasn't exactly the same (not sure what CB uses for font)... Take from it what you like.

The problem is some of those numbers don't line-up,and anyway,its been confirmed by multiple leaks the validation CPUs run at 2.8GHZ~3.2GHZ,and AMD had to adjust voltage manually to get their demo sample to run at 3.4GHZ it seems.

If you look at the AMD demos it hints at Broadwell level IPC(assuming SMT scaling is similar).

Now,people are argueing it looks much worse with these leaks - it really is not if you assume a sub 3GHZ base clockspeed,for these leaks,as it would point to Haswell to Broadwell level IPC.

Plus do we also really expect some random person on a forum,to be running a 3.5GHZ Ryzen CPU just barely a week after AMD could only demo a 3.4GHZ one??

AMD has not got Turboboost running properly yet,so what do you expect some bloke who got a random CPU in an early BIOS revision motherboard to do better??

 

[jpiniero]

AMD has not got Turboboost running properly yet,so what do you expect some bloke who got a random CPU in an early BIOS motherboard can achieve??

I find that hard to believe. If Ryzen is shipping in February, even a paper launch, they need to have started final production by now.

 

[Arachnotronic]

I find that hard to believe. If Ryzen is shipping in February, even a paper launch, they need to have started final production by now.

That's right. Takes three months for a wafer to go thru the fab, a little more time to package & test, then the chips need to be shipped out to OEMs/distributors.

The chip is done at this point if it's going to launch in Q1 2017.

 

[USER8000]

I find that hard to believe. If Ryzen is shipping in February, even a paper launch, they need to have started final production by now.

Well maybe you need to pay more attention to what was said during the presentation - the Ryzen chip had no Turboboost enabled,and AMD manually adjusted the voltage.

Multiple validation samples have been leaked at 2.8GHZ to 3.2GHZ clockspeeds.

So again please explain how some Chinese bloke on a forum suddenly got a fully functional motherboard with all the boost clockspeeds working properly,when AMD did not even manage to demo it at their big preview event??

How did this random person get the latest and greatest batch of chips from AMD??

Its most likely an early validation sample which has been leaked out on a motherboard with a less than final BIOS.

The scores make full sense if the CPU is running at a lowish clockspeed and actually ties in with the AMD demo.

Its the simplest explanation out there,instead of some weird microarchitecture which at one point rivals Broadwell and then is worse than Sandy Bridge. It looks a more traditional type of design,not the weird one which Bulldozer had.

Edit to post.

Do we even know how fast the RAM was running at either??

Those Intel HEDT Core i7 scores were quad channel DDR4 systems,and it is most likely Ryzen is more bandwidth limited being only dual channel for desktop systems,and some of the earlier AMD CPUs also responded well to lower timings.

Unless we have clockspeeds for both the CPU and RAM,we can't say whether the leaks are from a hobbled system or not even if they are genuine,and again we had situations when a buggy BIOS throttled chips - Gigabyte had that issue with its AM3+ motherboards.

 

[BeepBeep2]

You keep speaking of this however I and then Dresdenboy pointed out that based on the screenshot that sample has a base clockspeed ending in x.1 or x.4. There are two different points in the engineering sample name that identify the base clock speed and one that identifies boost clock.

If you are asking why the numbers don't physically line up because we didn't use the same font that is used in Cinebench.

Trust me, I'm sure there are Engineering Samples out there with clocks all over the place. Do you really think the only ones are 2.8 or 3.2 GHz base clock?

Nobody is claiming here that boost was enabled / disabled for this leak.

EDIT: Example
2D3101A2M88E4_35/31_N
__ ^ ^ ___________ ^ ^ ^ ^
Base_________Turbo/Base

E4 designates stepping

EDIT #2:
By the way RAM frequency / bandwidth does not affect Cinebench much.

 

[Doom2pro]

The problem is some of those numbers don't line-up,and anyway,its been confirmed by multiple leaks the validation CPUs run at 2.8GHZ~3.2GHZ,and AMD had to adjust voltage manually to get their demo sample to run at 3.4GHZ it seems.

If you look at the AMD demos it hints at Broadwell level IPC(assuming SMT scaling is similar).

Now,people are argueing it looks much worse with these leaks - it really is not if you assume a sub 3GHZ base clockspeed,for these leaks,as it would point to Haswell to Broadwell level IPC.

Plus do we also really expect some random person on a forum,to be running a 3.5GHZ Ryzen CPU just barely a week after AMD could only demo a 3.4GHZ one??

AMD has not got Turboboost running properly yet,so what do you expect some bloke who got a random CPU in an early BIOS revision motherboard to do better??

That 3.5 would be turbo boost, assuming it was off the sample was running at 3.1, a full 300Mhz less than the New Horizon sample.

 

[Doom2pro]

That's right. Takes three months for a wafer to go thru the fab, a little more time to package & test, then the chips need to be shipped out to OEMs/distributors.

The chip is done at this point if it's going to launch in Q1 2017.

You don't think last minute microcode and or metal layer changes can take place before mass production? Do you think AMD is mass producing Zen right now with final silicon?

 

[jihe]

Summit Ridge 8C/16T Cinebench R15 MT & Fritz Chess Scores Leaked (WCCFTech)

Comparison #1:

Core i7-7700K (Stock) 971* cb
Core i7-7700K (OC) 1108* cb
Core i7-6900K (Stock) 1565 cb
Core i7-6950X (Stock) 1863 cb
AMD RYZEN (8/16) 1188 cb

*http://www.bit-tech.net/hardware/cpus/2016/12/05/core-i7-7700k-performance-overclock-preview/3

Comparison #2:

Comparison #1:

Core i7-7700K (Stock)
35.52
Core i7-7700K (OC)
41.44
Xeon E5-2670 (Stock)
41.88
Core i7-6900K (Stock)
47.80
Core i7-6950X (Stock)
51.50
AMD RYZEN (8/16)
36.86

Core i7-6900K 30% faster @ FritzChess, 31% faster @ Cinebench R15.

http://wccftech.com/amd-ryzen-zen-cpu-benchmarks-leak/

Fritz score is pretty useless. For comparison my x5650@4.07Ghz scored 44.37 vs stock i7-6900k 47.80. LOL.
Basically Fritz score is simply proportional to cores x mhz, architectural improvements are completely ignored.

 

[Abwx]

An Athlon 845 Score about 318 - multiply that by 2 and we get:

636 for 8c + 40% = 890 for ZEN without SMT.

That doesnt work like this, at least when it comes to FP.

There s 2 FPUs in the Athlon 845 and there s 8 (improved...) FPUs in a 8C Zen, if a FPU score 159/3.5GHz then it should score 8 X 154 = 1232 in a 3.4GHz Zen, but that s without accounting any SMT gain or the fact that the FPU in a Zen core is quite bigger than in an XV module.

 

[USER8000]

You keep speaking of this however I and then Dresdenboy pointed out that based on the screenshot that sample has a base clockspeed ending in x.1 or x.4. There are two different points in the engineering sample name that identify the base clock speed and one that identifies boost clock.

If you are asking why the numbers don't physically line up because we didn't use the same font that is used in Cinebench.

Trust me, I'm sure there are Engineering Samples out there with clocks all over the place. Do you really think the only ones are 2.8 or 3.2 GHz base clock?

Nobody is claiming here that boost was enabled / disabled for this leak.

EDIT: Example
2D3101A2M88E4_35/31_N
__ ^ ^ ___________ ^ ^ ^ ^
Base_________Turbo/Base

E4 designates stepping

That 3.5 would be turbo boost, assuming it was off the sample was running at 3.1, a full 300Mhz less than the New Horizon sample.

Well if there were all these 3.5GHZ Zen chips floating about AMD would have benchmarked Zen running at 3.5GHZ by locking it at the maximum multiplier state,not 3.4GHZ,wouldn't they?

Also look at history,when AM3+ was launched ,Gigabyte AM3+ motherboards had throttling issues with FX CPUs,which meant they could not actually maintain their stated clockspeeds properly. Yet in Asus motherboards they were fine.

We have no clue what speed the memory is running at either - its an 8C chip running only dual channel memory and AMD has a recent history of worse memory controllers than Intel. I fully expect Ryzen to be bandwidth limited with so many cores just like the Phenom II X6 was,and I also fully expect that the extra bandwidth on socket 2011 will come into play with certain benchmarks.

AMD forced their demo chip to run at a constant 3.4GHZ,so it would not throttle,and I would expect them to have used a motherboard with the latest BIOS they could get their hands on.

This is why the leak is not really telling us anything new,because:
1.)We don't know how mature the BIOS on the motherboard is and whether the CPU is even hitting base clockspeed.
2.)We don't know how mature the BIOS on the motherboard is and what speed the RAM is being running at.

Plus we are also assuming the Chinese bloke didn't just use one of those 2.8GHZ validation samples and did some creative editing of the ID code.

If we were to believe what AMD has shown us,Ryzen is around Broadwell IPC,and those scores leaked would only really make sense at under 3GHZ anyway. After all HandBrake and Blender were pushing different loads,so I would not expect Cinebench to suddenly do worse,when it should be broadly covered by the previous benchmarks.

Even the FX8150 was only slightly worse(10% slower or around that) than a Core i7 2600k:

http://images.anandtech.com/graphs/graph4955/41691.png
http://images.anandtech.com/graphs/graph4955/41695.png

It would make little or no sense for AMD to suddenly do much worse in Cinebench compared to previous generational comparisons.

 

[BeepBeep2]

Well if there were all these 3.5GHZ Zen chips floating about AMD would have benchmarked Zen running at 3.5GHZ by locking it at the maximum multiplier state,not 3.4GHZ,wouldn't they?

Also look at history,when AM3+ was launched ,Gigabyte AM3+ motherboards had throttling issues with FX CPUs,which meant they could not actually maintain their stated clockspeeds properly. Yet in Asus motherboards they were fine.

We have no clue what speed the memory is running at either - its an 8C chip running only dual channel memory and AMD has a recent history of worse memory controllers than Intel. I fully expect Ryzen to be bandwidth limited with so many cores just like the Phenom II X6 was,and I also fully expect that the extra bandwidth on socket 2011 will come into play with certain benchmarks.

AMD forced their demo chip to run at a constant 3.4GHZ,so it would not throttle,and I would expect them to have used a motherboard with the latest BIOS they could get their hands on.

This is why the leak is not really telling us anything new,because:
1.)We don't know how mature the BIOS on the motherboard is and whether the CPU is even hitting base clockspeed.
2.)We don't know how mature the BIOS on the motherboard is and what speed the RAM is being running at.

Plus we are also assuming the Chinese bloke didn't just use one of those 2.8GHZ validation samples and did some creative editing of the ID code.

If we were to believe what AMD has shown us,Ryzen is around Broadwell IPC,and those scores leaked would only really make sense at under 3GHZ anyway. After all HandBrake and Blender were pushing different loads,so I would not expect Cinebench to suddenly do worse,when it should be broadly covered by the previous benchmarks.

Even the FX8150 was only slightly worse(10% slower or around that) than a Core i7 2600k:

http://images.anandtech.com/graphs/graph4955/41691.png
http://images.anandtech.com/graphs/graph4955/41695.png

It would make little or no sense for AMD to suddenly do much worse in Cinebench compared to previous generational comparisons.

The vcore increase you speak about is related to a short bitsandchips rumor that also surfaced today naming no sources and the reason stated was not to stop it from throttling, but to ensure the system was stable.

You don't need to explain anything about AMD's desktop plaforms to me, USER8000. I know every AMD AM socket platform from a user / overclocker perspective better than I know the back of my own hand.

And no, I'm sure the ES used for New Horizon had a boost clock designated to it too, they just disabled boost for the event.

Once again RAM speed does not affect Cinebench much at all.

 

[Vesku]

I would not be surprised if Zen IPC varies from SB/IB to Haswell/Broadwell depending on task. Have to assume AMD is picking some of the best workloads to justify their planned pricing. Given they are trying to show it as competitive to the 6900k makes me think the full 8 cores will be priced closer to $500 than $200.

 

[rvborgh]

Any possibility of this being the 6C/12T (basically 2 cores disabled) variety of Zen? 0.75 * 1592 ends up being almost right on the money. Is it known that the ES had all the cores working?

 

[Abwx]

It would make little or no sense for AMD to suddenly do much worse in Cinebench compared to previous generational comparisons.

It s possible because Cinebench R15 benchmark has been tailored to give Intel more advantage than in CB 11.5, otherwise they wouldnt had re sized the scene if it was only about an updated renderer.

From CB 11.5 to CB R15 the FX8350 went from 6.45% advantage over the i7 2600K to a 7% disadvantage, that s 14% "IPC" gained from nowhere, at this rate they got more thanks to those shenanigans than to uarch evolutions...

Indeed CB R15 produce results that are at odd with Povray or even 3ds Max, in both those similar apps the 2600K is left in the dust..

There s two renderings using two different engines within 3DS MAX in Hardware.fr reviews :

http://www.hardware.fr/articles/946-5/performances-applicatives.html

Povray scores can be checked at Computerbase.de :

https://www.computerbase.de/2015-08...0k-test-benchmark-skylake/6/#diagramm-pov-ray

https://www.computerbase.de/2016-05/intel-core-i7-6950x-6800k-test/4/#diagramm-pov-ray

 

[BeepBeep2]

Any possibility of this being the 6C/12T (basically 2 cores disabled) variety of Zen? 0.75 * 1592 ends up being almost right on the money. Is it known that the ES had all the cores working?

If it is even close to the CPU we think it is based on the ES OPN, it is an 8-core CPU

The result is still almost Haswell-E level IPC, so Dresdenboy (and I) am not sure if it is fake or not. The part OPN suggests not, but anyone knowing how AMD names their ES parts could theoretically fake that.

Assuming Summit Ridge @ 3.1 GHz clock and not in a boost state (this chip shows 3.5 GHz boost...) and 1188 points, you get around 1350 points @ 3.5 GHz
Haswell-E 8c/16t 5960X @ 3.0 Base and 3.5 GHz Turbo (not sure of boost clock when under Cinebench load, maybe 3.3 GHz?) : ~1350 points

Could be plausible. Broadwell-E does exceptionally well in Cinebench R15, about 10% uplift clock-per-clock over Haswell-E.

 

[rvborgh]

Hi Abwx,

When i run CB on my Intel machines (i7-4770/i7-5930K) and compare the scores to my 48 core K10 Opteron rig... usually the range of K10 cores per Intel core ends up being around 2.5 for the same performance per clock - therefore i pretty much "price" things in K10 units since i've got an abundance of them. We also saw on Blender that each Zen core pushes at least 2.5 K10 cores (when comparing my Opteron scores).

BD arch is just inherently terrible at CB due to not having an FPU per core... each PD core typically is only worth .85 of a K10 core at same clock on CB.

K10 for what it is... runs CB quite well (my rig scores 39.04 after all at 3.1 GHz all cores, or 16 when i run it tied to 16 cores running at 3.6 GHz) and the scaling is close to perfect (although you have to have node interleaving enabled when running it to achieve this) all the way out to 48 cores.

If 8C Zen is performing on the level of Haswell in CB... then it should ballpark perform the same as 20 K10s. i end up with a score of between 1484 and 1592 in R15 when "priced" that way at 3.4 GHz.

A score of 1188 would mean the loss of 5 K10 units of computing power. i just cannot see that given how well CB scales. However it would match up with a 6C/12T Zen almost perfectly. The other possibility would be drastically reduced GHz (8C/16T between 2.5 and 2.6 GHz? Did the ES run that slow? Or some other malady).

It s possible because Cinebench R15 benchmark has been tailored to give Intel more advantage than in CB 11.5, otherwise they wouldnt had re sized the scene if it was only about an updated renderer.

From CB 11.5 to CB R15 the FX8350 went from 6.45% advantage over the i7 2600K to a 7% disadvantage, that s 14% "IPC" gained from nowhere, at this rate they got more thanks to those shenanigans than to uarch evolutions...

Indeed CB R15 produce results that are at odd with Povray or even 3ds Max, in both those similar apps the 2600K is left in the dust..

There s two renderings using two different engines within 3DS MAX in Hardware.fr reviews :

http://www.hardware.fr/articles/946-5/performances-applicatives.html

Povray scores can be checked at Computerbase.de :

https://www.computerbase.de/2015-08...0k-test-benchmark-skylake/6/#diagramm-pov-ray

https://www.computerbase.de/2016-05/intel-core-i7-6950x-6800k-test/4/#diagramm-pov-ray

 

[rvborgh]

PS: evidently some engineer is posting here that they are already receiving actual Zen chips (my Mandarin is admittedly a bit rusty so i may be mistranslating) instead of ES... and so they can actually test the real thing. Can a native Chinese reader comment on this post and confirm this?

http://cmt.pconline.com.cn/topic/a0/r0/p7/ps30/t13500173.html

 

[StrangerGuy]

A score of 1188 would mean the loss of 5 K10 units of computing power. i just cannot see that given how well CB scales. However it would match up with a 6C/12T Zen almost perfectly. The other possibility would be drastically reduced GHz (8C/16T between 2.5 and 2.6 GHz? Did the ES run that slow? Or some other malady).

1188 at 16T + 3.1GHz isnt shabby at all, which is IPC between IB and HW and it will be ~60-70% faster in MT than a 4790K with a 4.2GHz OC. Heck, if you want a 16T Zen for <$300 that's what you want to hope for.

 

[Abwx]

BD arch is just inherently terrible at CB due to not having an FPU per core... each PD core typically is only worth .85 of a K10 core at same clock on CB.

K10 for what it is... runs CB quite well (my rig scores 39.04 after all at 3.1 GHz all cores, or 16 when i run it tied to 16 cores running at 3.6 GHz) and the scaling is close to perfect (although you have to have node interleaving enabled when running it to achieve this) all the way out to 48 cores.

Given the numbers it s obvious that you re talking of CB 11.5 while i was questioning CB R15 that saw a 12% regression on AMD s CPUs compared to Intel s, and wich seems to compress the improvements made with Steamroller and Excavator.

See the pattern.?..

The Excavator based Athlon 845 is 21% faster than the i3 4130T in CB 11.5, but in CB R15 the advantage is reduced to 7%, yet there are people using CB R15 as thermometer to allegedly measure IPC.

For the fun, a 64bit precision renderer that make use of all FP instructions :

If 8C Zen is performing on the level of Haswell in CB... then it should ballpark perform the same as 20 K10s. i end up with a score of between 1484 and 1592 in R15 when "priced" that way at 3.4 GHz.
.

Your estimation is accurate methink but we ll have to wait for actual tests, at first look i would say that Zen could match the 6900K in CB 11.5 but it s unlikely in CB R15 given the behaviour of this "benchmark"...

 

[jhu]

For the fun, a 64bit precision renderer that make use of all FP instructions :

If by "all FP instructions" you mean "nothing beyond scalar SSE2", then sure. Otherwise it wouldn't run on the Pentium G3258. Povray doesn't do code dispatching.