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

[krumme]

< 297mm² nominal for all Orochi dies.

Ok. A tiny bit less bloated

 

[sm625]

http://www.pcper.com/news/Processors/Leaked-Geekbench-Results-AMD-Zen-Performance-Extrapolated

Welp. I guess that explains why AMD stock is tanking. Zen is literally going to be DOA. No one is going to buy a 3.2GHz Zen that performs the same as a 2.5GHz 3570K. If Zen cant even hit 3.2GHz it will be even worse. KBL Core m could very well beat a 95W Zen in single threaded tasks.

 

[lolfail9001]

It was a 315mm2 cpu and from the viewpoint of winzip at least it was very much a 8 core chip. The same way this blender test exposes all the best of the fpu in zen.

But we know its not going to use a gazilion watts and be 315mm2. On the contrary we have a 32c 180w underway.

The point is the situaion is in no way comparable to then. But ofcource those hoping for a skylake desktop competitor on st is ofcource looking the wrong way.

From the viewpoint of winzip it was a weird 8 core.
From the viewpoint of anything fpu-heavy, it was a quad core with more ALUs.

Anyways, only few months to go.

 

[DrMrLordX]

I have really bad memory in regards to things that I just simply glance over. Where was the total rebuild guide for Unix operating systems for OpenCL-HSA or HSA-Compiler instead of C/C++?

Most of what you want (if it exists) that is not deprecated is here:

https://github.com/hsafoundation

You'll have to drill down to find exactly what you want. It looks like most of the "old" stuff is deprecated and has been replaced by Radeon Open Compute (ROCm), so . . . that's all I got right now.

For OpenCL->HSA there's always CLOC which still exists (may not be what you want, but I'll link it anyway):

https://github.com/HSAFoundation/CLOC

It looks like it's one of the things receiving regular updates.

The HSA-enabled LLVM dev project ist kaput, and the HSA-enabled GCC branch hasn't seen an update since July (which actually isn't too bad).

From the viewpoint of winzip it was a weird 8 core.
From the viewpoint of anything fpu-heavy, it was a quad core with more ALUs.

Anyways, only few months to go.

Fortunately, the server/workstation/HEDT world(s) do not live and die by Geekbench.

 

[KTE]

But we know its not going to use a gazilion watts and be 315mm2. On the contrary we have a 32c 180w underway.

We don't know any of this yet (power).

Or if it's clocks suck like Barcelona.

Or if it's performance is 5 years behind Intels.

No, we really don't. Except the die size.

If it is not competitive, they really have no chance. The market is already 98% Intel, and only loves to look at Intel.

AMD has a very tough uphill struggle to even convince companies to consider them today (except HPC).

However, if this chip tanks on mobile, desktop, workstation and most of server -- it could still find a niche. Like in Cloud.

Speaking as an Infrastructure worker in one of the biggest corps in the world -- every major corp I've liaised with is now buying IaaS/PaaS. Gone are the days of locally hosted. You buy VMs from datacentres who sell licensed by the Core. They absolutely love these many low-speed cores. For the hosters, this is what sells. As extra Cores can be included but left on "standby"... Only used and paid for IF required at any one time.


Sent from HTC 10
(Opinions are own)

 

[krumme]

Fortunately, the server/workstation/HEDT world(s) do not live and die by Geekbench.

And neither do Intel if the comparison is A10. But i guess for zen it will do. Lol.

I would personally be very happy if zen is so stong it can move price of higher end desktop 4c core or 6c low end eg 6800. The last years have been so boring even a small change is welcome. It would be nice if we could move to Intel core 6c 12t. From the usual 4c setup.

 

[krumme]

We don't know any of this yet (power).

Or if it's clocks suck like Barcelona.

Or if it's performance is 5 years behind Intels.

No, we really don't. Except the die size.

If it is not competitive, they really have no chance. The market is already 98% Intel, and only loves to look at Intel.

AMD has a very tough uphill struggle to even convince companies to consider them today (except HPC).

However, if this chip tanks on mobile, desktop, workstation and most of server -- it could still find a niche. Like in Cloud.

Speaking as an Infrastructure worker in one of the biggest corps in the world -- every major corp I've liaised with is now buying IaaS/PaaS. Gone are the days of locally hosted. You buy VMs from datacentres who sell licensed by the Core. They absolutely love these many low-speed cores. For the hosters, this is what sells. As extra Cores can be included but left on "standby"... Only used and paid for IF required at any one time.


Sent from HTC 10
(Opinions are own)

I asume this a cloud cpu.
And made for that market. As the primary segment bar none.

As i can tell thats the future because how software is driven, even far more than it is now, and i am personally betting a lot on it. I think we have just seen the start of it.

For that it can have "barcelone freq" so to speak but if it fail on efficiency its down the drain.

We have a lot of ivy bridge xeon out there that need some cheap fast efficient opponent to get changed. It takes something for that to be good tco. Its a tough job.

 

[Dresdenboy]

Welp. I guess that explains why AMD stock is tanking. Zen is literally going to be DOA. No one is going to buy a 3.2GHz Zen that performs the same as a 2.5GHz 3570K. If Zen cant even hit 3.2GHz it will be even worse. KBL Core m could very well beat a 95W Zen in single threaded tasks.

So you think, the GloFo amendment and stock + senior notes offer and the tanking markets didn't weigh as much as this benchmarked not so final ES?

With such a proof of failure, it would be wise to sell everything except the family and take the money as margin for a big short on $AMD.

 

[frozentundra123456]

Not to mention a normal reaction to a huge run up in the last few months.

 

[lolfail9001]

Fortunately, the server/workstation/HEDT world(s) do not live and die by Geekbench.

You are right, but that is boring way around it.

 

[coercitiv]

that does not mean that any version of smt gives you gains. (other then being able to run 2 threads at once)

So it is your opinion that the SMT implementation in Zen will offer no performance gains in MT scenarios?

 

[blublub]

Welp. I guess that explains why AMD stock is tanking. Zen is literally going to be DOA. No one is going to buy a 3.2GHz Zen that performs the same as a 2.5GHz 3570K. If Zen cant even hit 3.2GHz it will be even worse. KBL Core m could very well beat a 95W Zen in single threaded tasks.

Yeah and the share dilusion has nothing to do with the stock going down to the ammount it is worth with new total numbers of shares.. ...all these traders just cant do math they just look at leaked benches....man wake up!

Additionally please remebrr Skylake ES leaks with frequency of only 2.8Ghz - aLOT can change until release.

 

[KTE]

Talking Cloud hosting... VMs...

AMD already has a Warsaw 16-core server CPU based on Bulldozer 2nd Gen (Piledriver) 2.3GHz base, 2.6GHz full load and 3.2GHz half-load... at 115W, in the market since 2012.

Nothing with SR/XV, but assuming it did not originally breach 115W under load, down to 85-95W envelope would be a given after 4 years of tuning and maturity (even PD to SR in 2 years became 95>65W -- remember TDPs are not actual power draw).

Moving that to 14nm FF, I'd expect, away from the obvious size shrinkage, the same XV to run at 2.6GHz base at 85W as the minimum.

With the FO4 being similar, wider + increased resources + SMT, Zen on the same process should enable (+15% power) a 2.3GHz base 16-core/32-thread CPU at 105-115W still on 28nm (GF28A).

Combined with a full node shrink + FF (-25% power), I'd expect 2.3GHz base for a 18-core 36-thread 115W CPU, as the very minimum if the process was OK. Up to 2.6GHz I'd expect as a given.

 

[The Stilt]

Talking Cloud hosting... VMs...

AMD already has a Warsaw 16-core server CPU based on Bulldozer 2nd Gen (Piledriver) 2.3GHz base, 2.6GHz full load and 3.2GHz half-load... at 115W, in the market since 2012.

Nothing with SR/XV, but assuming it did not originally breach 115W under load, down to 85-95W envelope would be a given after 4 years of tuning and maturity (even PD to SR in 2 years became 95>65W -- remember TDPs are not actual power draw).

Moving that to 14nm FF, I'd expect, away from the obvious size shrinkage, the same XV to run at 2.6GHz base at 85W as the minimum.

With the FO4 being similar, wider + increased resources + SMT, Zen on the same process should enable (+15% power) a 2.3GHz base 16-core/32-thread CPU at 105-115W still on 28nm (GF28A).

Combined with a full node shrink + FF (-25% power), I'd expect 2.3GHz base for a 18-core 36-thread 115W CPU, as the very minimum if the process was OK. Up to 2.6GHz I'd expect as a given.

Except on Opterons the TDP applies only on the base (non-boosted) frequency and is a "typical figure" (don't know the actual definition for this), not the maximum unlike on consumer parts.

Opteron 6380 - 115W TDP - 2.5GHz base - 133.2W maximum power @ base
Opteron 6386 - 140W TDP - 2.8GHz base - 156.7W maximum power @ base

 

[jpiniero]

Talking Cloud hosting... VMs...

I don't know if that kind of estimation is going to work. You have to remember that there's only one 8-core die involved but you can get up to 4 of them on a socket connected via GMI. Even Warsaw is basically a 2 die model.

The one die base is 2.8 Ghz at 95 W.
The four die base is 1.44 Ghz at 180W (?) (45 W/die, seems like it should be less with the clock speed halved)

 

[The Stilt]

I don't know if that kind of estimation is going to work. You have to remember that there's only one 8-core die involved but you can get up to 4 of them on a socket connected via GMI. Even Warsaw is basically a 2 die model.

The one die base is 2.8 Ghz at 95 W.
The four die base is 1.44 Ghz at 180W (?) (45 W/die, seems like it should be less with the clock speed halved)

You cannot expect linear scaling for the core clock, when each die has much more than just the CPU cores. In MCM configuration some parts of individual dies become unnecessary, so that will also bias the clocks somewhat.

 

[KTE]

Except on Opterons the TDP applies only on the base (non-boosted) frequency and is a "typical figure" (don't know the actual definition for this), not the maximum unlike on consumer parts.

Opteron 6380 - 115W TDP - 2.5GHz base - 133.2W maximum power @ base
Opteron 6386 - 140W TDP - 2.8GHz base - 156.7W maximum power @ base

That's why I'm talking server and kept it inline with the official TDPs.

Frequency curves aren't linear and each process/processor scales differently with voltage across the curve. There is always a range the process is optimized for. But with any one CPU, across full node shrinks, the gains are typically >20% lower power and >15% higher switching frequency.

6386 is a non-standard PD SKU, SE. It's the top one that sacrifices power for maximum speed. It also hits 3.2GHz >8core load.

Also the reason for the TDP figure used as such is, with that many cores, full pipeline utilization for such power draw is rare and difficult. Usually the CPU is bottlenecked elsewhere far before that is possible.

The one die base is 2.8 Ghz at 95 W.
The four die base is 1.44 Ghz at 180W (?) (45 W/die, seems like it should be less with the clock speed halved)

The two quoted CPUs are Zen based? I think you are talking about different dies here, 1 Server and 1 DT.

180W sounds inline for the actual maximum power for an 18-core 14nm, but not for 1.44GHz.

That's Barcelona 65nm again.

Sent from HTC 10
(Opinions are own)

 

[The Stilt]

I think you are talking about different dies here, 1 Server and 1 DT.

There is only a single Zeppelin die. Server parts will use multiple of them (2 - 4).

 

[Abwx]

The one die base is 2.8 Ghz at 95 W.
The four die base is 1.44 Ghz at 180W (?) (45 W/die, seems like it should be less with the clock speed halved)

It s within 95W for a 8C/16T at 3GHz according to AMD, then it will be also about 95W for 32C/64T but at 1.5GHz, that s what laws of physics say, and still, that s assuming that at 3GHz transistors are still in a favourable part of their curves wich i assume is the case, if they had to overscale the voltage to get at this frequency and are sill within 95W/8C then it s less than 95W for the 32C/1.5GHz case.

 

[Glo.]

Abwx, its highly unrealistic. IIRC the TDPs, that appeared somewhere for the server parts are 150 and 180W. Secondly, do not take seriously the core clocks that you see right now, they will be higher on all of the CPUs from range, at least thats what logic dictates.

Thirdly. 95W TDP IMO is only for "modest" mid-range High-End desktop CPUs from AMD. Wraith Cooler is designed to work with 125W TDP, so I think we have to look at 8 core Black Edition Zen CPU at that TDP, with 4 and 6 core variants at 95W.

 

[Abwx]

Abwx, its highly unrealistic. IIRC the TDPs, that appeared somewhere for the server parts are 150 and 180W. Secondly, do not take seriously the core clocks that you see right now, they will be higher on all of the CPUs from range, at least thats what logic dictates.

Thirdly. 95W TDP IMO is only for "modest" mid-range High-End desktop CPUs from AMD. Wraith Cooler is designed to work with 125W TDP, so I think we have to look at 8 core Black Edition Zen CPU at that TDP, with 4 and 6 core variants at 95W.

The 95W are for the stated 32C/1.5GHz, the same logic applied to 150-180W would yield 1.88-2.06GHz.

Now if the clocks at 95W are higher than 3GHz as suggested by AMD the frequencies above should be revised accordingly and on a linear faschion, FI if it s 3.3GHz/95W then the frequencies i quoted shoud be increased by 10%, at same TDPs of course.

So far the indications given by AMD point to better efficency at 3GHz than BDW, this imply that the CPU is not constrained by any limitation in the vicinity of this frequency and that it should clock substancially higher at still reasonable efficency.

 

[Glo.]

I don't think that core clocks, and power consumption scale that linearly as you point it out.

 

[Abwx]

I don't think that core clocks, and power consumption scale that linearly as you point it out.

Because you thought that the numbers above are linear scalings..?.

These are power scaling wich mean that power increase as a square of frequency, most ironic is that some people pointed that it could be a scaling that is far from ideal, ideal being namely this square law, but then if they are above a square law, and close to a cubic law, then it means that from 3GHz to 1.5GHz, or whetever lower frequency, power would scale down much more than what i posted above...

 

[lolfail9001]

Because you thought that the numbers above are linear scalings..?.

These are power scaling wich mean that power increase as a square of frequency, most ironic is that some people pointed that it could be a scaling that is far from ideal, ideal being namely this square law, but then if they are above a square law, and close to a cubic law, then it means that from 3GHz to 1.5GHz, or whetever lower frequency, power would scale down much more than what i posted above...

Yeah, but you ignore the entire fact that Naples has all the uncore (that consumes relatively constant amount of power) of 4 ZP dies.
Account for that next time.
And finally, you ignore the fact that power-frequency depends on voltage-frequency you have no clue of right now.

 

[superstition]

A huge die as a bonus.

Bd was 315 mm2 !!! Good grief.

Not seeing a vast difference, especially with The Stilt's < 297mm² figure.