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

[cbn]

Looking at the Zen graph in this post it looks like the Zen Core at the point of the tachometer corresponds to around 12.5W.....but notice it is not leveling off at this point.

So maybe there is some decent room left in the process and design.

8 x 12.5W = 100W. Add in another 15W for the integrated chipset and maybe TDP would be 115W. Assuming extra room is indeed left then I think 125W+ could be possible for 8C/16T. With that mentioned, I am mostly concerned with the 4C/8T. Considering AMD already has nice copper cored 95W coolers available it would be good to at least see a cTDPup option (maybe 95W, if not then 80W) on the 65W SKUs.

It doesn't work like that, even remotely.

Two Excavator CUs require > 45W (other SoC HW excluded) to sustain 3.4GHz during Cinebench R15, and that's not even the most stressful real world workload you can throw at it.
Not to mention that the Fmax vs. Vdd curve on 14nm LPP is vastly steeper than on 28nm HPP, close and beyond the optimal Fmax range. 14nm LPP is superior to 28nm HPP when it operates at it's optimal range, but anything outside this range is far inferior.

Hi Stilt,

I was just speculating about the max TDP for 8C/16T. Not sure what the clocks would work out to be ......but if I had to guess such a part would have little in the way of turbo (ie, higher base clock, but much smaller turbo)

Also, perhaps AMD would not release such a part (in the beginning) because they don't have enough dies that could scale all 8C to the same level high clock/high TDP per core?

 

[VirtualLarry]

There is 2x 10GBe on die, but honestly that is the first part I'd fuse off in a home cpu to be able to ask for more for the server ones.

Are you kidding? Crippling "platform features" isn't good for the enthusiast market.

If it's already in the silicon, why not offer a +$50 SKU option with 10GbE enabled (just one)?

I think that they would gain more from enthusiast sales and marketshare, than they would lose from smaller server customers buying enthusiast platforms rather than true server boxes. (Who builds their own servers? Only the very small, and very largest customers. Most get from resellers like Dell and HP.)

Edit: Lord knows the SOHO business workstation / home server market needs some client platform 10GbE options to kick 10GbE into the mainstream, for home LANs. Zen could spark that renaissance.

 

[itsmydamnation]

licencing fee's could play a big part amd didn't design the 10gbe but you think $50 usd would be more then enough to cover that. They could just expose one as well.

 

[cbn]

Are you kidding? Crippling "platform features" isn't good for the enthusiast market.

If it's already in the silicon, why not offer a +$50 SKU option with 10GbE enabled (just one)?

I think that they would gain more from enthusiast sales and marketshare, than they would lose from smaller server customers buying enthusiast platforms rather than true server boxes. (Who builds their own servers? Only the very small, and very largest customers. Most get from resellers like Dell and HP.)

Edit: Lord knows the SOHO business workstation / home server market needs some client platform 10GbE options to kick 10GbE into the mainstream, for home LANs. Zen could spark that renaissance.

With budget PCIe 3.0 x 2 M.2 SSDs on the way (eg, Marvell 88NV1160 controller) capable of 1600 MB/s read speed having that 10 GbE (1250 MB/s) could be interesting.

Also, another thing to consider would be game streaming.

 

[Dresdenboy]

In those Zen related slides there s this one :

So Finfets used for Zen not only consume less but also have higher working frequency in respect of a 28nm that can only be the one used for Kaveri/Carrizo/Bristol Ridge.

This graph looks like the one used for Polaris:

 

[Dresdenboy]

Took me a while to realize something from what AMD stated:

How much does dual channel vs. quad channel hurt Broadwell-E? (i.e "comparable configuration", Zen has just dual channel memory).

Good point. But renderers running such simple scenes with some locality for all threads usually run from cache.

Even F. Piednoel, who seemingly wasn't that happy on Twitter about the demo, mentioned Blender loops running just from L1 in his tests. But he complained, that AMD would use cache hierarchy information, which would render (pun intended) IPC comparisons useless. But the same happens with BDW, Willamette, Katmai, etc.

An "AMD"-codepath related complaint was moot, because Blender has a feature called "Approximate Minimum Degree".

 

[Hail The Brain Slug]

Good point. But renderers running such simple scenes with some locality for all threads usually run from cache.

Even F. Piednoel, who seemingly wasn't that happy on Twitter about the demo, mentioned Blender loops running just from L1 in his tests. But he complained, that AMD would use cache hierarchy information, which would render (pun intended) IPC comparisons useless. But the same happens with BDW, Willamette, Katmai, etc.

An "AMD"-codepath related complaint was moot, because Blender has a feature called "Approximate Minimum Degree".

I have a hunch that AMD's chosen blender benchmark was deliberate to mask the areas where they still may not be competitive with Intel's E series processors. I did some testing with Blenchmark (The closest thing I could find) to see if I could pinpoint why they may have chosen Blender specifically for their demo.

I ran a series of tests with my 5960x to see the affect cache speed and memory bandwidth played in the benchmark performance. To sum it up, I ended up taking out 3/4 of my memory sticks to force single channel and dropped my cache clocks down to 2.75 GHz and the benchmark performed identically to full speed cache/quad channel. (This is with core at 4.5 GHz, where it should have been more sensitive to a cache or memory bottleneck). The only time I got performance to drop was downclocking the CPU core speed.

Considering historically AMD's cache performance has lagged behind intel and the fact that Zen is a dual channel platform, I sort of get the idea they chose Blender to hide both of those shortcomings when comparing to a BW-E processor. I think when we start seeing benchmarks of programs that are more sensitive to these two things, we will see Zen start to fall behind.

 

[cbn]

In those Zen related slides there s this one :

So Finfets used for Zen not only consume less but also have higher working frequency in respect of a 28nm that can only be the one used for Kaveri/Carrizo/Bristol Ridge.

This graph looks like the one used for Polaris:

It looks almost exactly the same. The only difference I notice is that the top graph has a little upturn at the very end for the blue dashed lines.

 

[Erenhardt]

I don't think Zen will ship with much higher clocks. It they can get anywhere close broadwell-E at 95 Watts, they made a huge leap in terms of ipc and efficiency.

Asking for more than 3.2 base clock is wishful thinking and setting for a disappointment. 95 watts is only so much. And every integrated component is cutting it thinner.

That said, I think the limits is much higher when you lift TDP limit, i.e. overclocking should be great if you can feed enough amps and motherboard manufacturers provide hardware and software support for over-specification CPU settings.

I hope amd sticks to their FX line history and all Zen CPUs are unlocked.
Also am4 motherboard better be good on power delivery Those Zen cores will need a lot of juice at higher frequencies and it would be nice if APUs maintained their turbo cores at GPU loads.

 

[William Gaatjes]

I was doing some background reading on AMD. I was thinking and reading about the jaguar cores. These were designed to be easy synthesizable and portable between foundries. It makes sense for a cash strapped AMD to continue down that road.
What if the new ES zen core we see now also is designed to be easily synthesizable and portable ?
Assuming AMD would win future console wins, jaguar cores will not cut it in the future. But Zen cores at clock frequencies a bit lower than now presented would.

On a side note :
Do you guys think that polaris is also designed in a similar fashion ? To be portable ?
Perhaps Vega is the first true new optimized design for AMD. Not to be easily synthesizable and portable.

 

[aegisofrime]

If it brings more affordable 4/8 cpu or making it the new midrange , no matter which side you buy , zen is a win for me .

I'm not comfortable with the assumption that Zen will be a lot more afforable than Intel offerings. If you guys were around during the Athlon XP and Athlon 64 era, AMD charged according for them. AMD is not running a charity.

I see the same mistaken assumption before the launch of RX 480. People were actually expecting it to cost $199 and beat a GTX 1070. If it were capable of beating a 1070 it would be priced accordingly. AMD charges what they can, people. Let's not make that mistake.

 

[DrMrLordX]

Not saying that AMD did such thing, but so far nothing AMD has said (40% over Excavator) implies that they would get anywhere close (within ~20%) to Broadwell / Skylake IPC.

It would if when they said IPC, they also meant throughput.

 

[The Stilt]

It would if when they said IPC, they also meant throughput.

They exactly said IPC.

 

[TheELF]

They exactly said IPC.

IPC was coined when there where still CPUs(central processing unit) which one is the central processing unit today? We have 4-8-16 CPUs inside our CPUs nowadays.
When people on this forum say IPC of ivy or haswell they sure aren't talking about "total IPC of all cores" since something like that is easily achieved by stringing together a lot of slow cores, they mean single core performance at the same clocks.

AMD helps this along by being very elusive on the concept of "core" I mean they sell 10 and 12 core laptops because they decided that GPGPUs are totally equal to CPUs,they called the FX construct a module running two threads,on this point, FX had CMT,clustered multithreading, because each thread had not the same amount of resources available at all times,so how would you call this thing if each "thread" (core) had the same amount of resources available at all times?
It may very well be that the 8c/16t zen will have 4 quad core modules,so in total 16 of the things we would call cores,effectively cutting "IPC" (of a core) in half.

 

[ioni]

I see the same mistaken assumption before the launch of RX 480. People were actually expecting it to cost $199 and beat a GTX 1070. If it were capable of beating a 1070 it would be priced accordingly. AMD charges what they can, people. Let's not make that mistake.

What? No one expected it to beat a 1070. Or even come close to it. Especially since it was being advertised at half the price. People did expect it to be $199 though.

It's perfectly reasonable to think that a competitive AMD chip will result in lower CPU prices. But the 480 was never meant, nor advertised, to compete with the 1070.

 

[Abwx]

This graph looks like the one used for Polaris:

Of course, if a Finfet switch faster in a given circuitry that was previously using 28nm it will do the same in any other previous 28nm circuitry..

 

[frozentundra123456]

Aside from the grain of salt necessary for any vendor supplied benchmark, the fact that AMD is comparing Zen to a $1000 octo core Broadwell-E processor (favourably at that), and not a quad core Skylake, already places Zen far above my original expectations.

Assuming that Zen ends up at least somewhat competitive against Intel's best in reality as well, the big question is whether Intel and AMD will cooperate in keeping each other's margins high, or if we'll see a return to the competition of old and these $1000+ processors dropping to a more appropriate price. It would sure be nice if the answers were yes, and the later.

I remember AMD also compared Bulldozer to a thousand dollar intel extreme edition cpu before launch in a gpu limited gaming scenario, and at that time the extreme edition was not even the best gaming cpu because of lower stock clocks. Of course the "takeaway" from AMD was that Bulldozer was equivalent to a thousand dollar cpu.

 

[blublub]

Dies quad vs Dual channel really matter in real world scenarios

 

[jihe]

Who are these AMD marketing people? The Zen logo looks like a gigantic piece of turd already!

 

[R0H1T]

Dies quad vs Dual channel really matter in real world scenarios

No it doesn't, only a handful of consumer apps benefit from quad channel memory. I doubt they're even in the double digits (the more popular ones) & even then high speed DDR4 in dual channel is much preferred because of stability & the fact that there aren't too many CPU's which can handle 4GHz quad channel memory, or make better use of it.

 

[imported_bman]

Who are these AMD marketing people? The Zen logo looks like a gigantic piece of turd already!

Zen hence the Enso.

 

[Dresdenboy]

Who are these AMD marketing people? The Zen logo looks like a gigantic piece of turd already!

I've seen the Zen circle at local shops already. But it might look like sth. different to people with a predefined opinion.

 

[sirmo]

Beating Broadwell-E clock for clock even if it's just Blender is still an amazing achievement. To design an architecture from scratch and leapfrog decades of Intel's CPU development (on an inferior process no less, with a fraction of the R&D budget) is nothing short of amazing. Many of us hoped for near Haswell IPC. Remains to be seen how it clocks and how well it does in other disciplines, but this is looking really promising.

 

[blublub]

Beating Broadwell-E clock for clock even if it's just Blender is still an amazing achievement. To design an architecture from scratch and leapfrog decades of Intel's CPU development (on an inferior process no less, with a fraction of the R&D budget) is nothing short of amazing. Many of us hoped for near Haswell IPC. Remains to be seen how it clocks and how well it does in other disciplines, but this is looking really promising.

Even if it bottoms out at 3 or 3.1 GHz for the 8c it will still be enough given the 95w TDP to sell good.

 

[IEC]

Personally interested in seeing what the Fmax will be. If max OC Zen is comparable to max OC Broadwell-E then I will buy it. Otherwise, I would consider Skylake-E instead. I don't care about power usage.