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

[realibrad]

Power does not depend solely on the CPU. We don't know what motherboard and other stuff AMD put in the test config, because >100W idle is ridiulcously high.

As a matter of fact, any modern CPU should dissipate close to 0W at idle. It gets power gated. Or do you think laptops also consume 100W at idle? I'm quite sure my laptop has about ~10 hours of battery life (which also includes the screen power consumption) when I'm just browing.

So idle CPU power consumption is 0W. Then we just have to look at how much power is added when the system is in load because that power will be coming from the CPU.

AnandTech also uses power delta, so it just shows how biased people are if you dismiss this calculation.

But here are some idle power consumptions:

The last Intel CPU to get >100W idle is Ivy Bridge.

So AMD probably attached a 50W light bulb to the Broadwell system to get similar idle powers so they wouldn't get humiliated?

What does that have to do with your statement of

So tell me, who has the highest performance per watt?

You clearly answered your own question when you posted this.
Zen idle: 93W
Intel idle: 106W

Zen load: 187W
Intel load: 191W


If zen is more efficient at idle, and more efficient at load, then what is your point?

Assuming your numbers are true, Zen uses 12% less at idle, and 2% less at load.

By your delta logic, a cpu that has 200W idle, and 200W load is better than a cpu with 100W idle and 180W load because the delta for the 2nd is 80W whereas the first is 0. You do not understand numbers if you think that argument is relevant.

Also, did Intel attach a 50W light bulb to the 6900k to make the 6600k look better? I mean, the power usage of the 6900k is so high.


If a CPU uses less power and idle, and less power at load, and has equal or better performance, then its more efficient. Simple math.

 

[superstition]

Currently they've only shown 8Cores+SMT favouring tests. MOAR CORES marketing.

I don't know. My 2009 4 core/thread i5 750 at 3.8 Ghz beat someone's 4 GHz 8 core/thread 8350 with the 100 samples test.

That suggests to me that either his CPU is throttling or the test is mostly of FPU performance.

update:

Someone else's posted time for 200 samples on a 4.6 GHz FX is basically identical to my Lynnfield's performance at 3.8. This seems to confirm my belief that this test is not taking advantage of 8 integer cores.

 

[bjt2]

Its easy to manipulate an experiment for a desired result.

Hence why independent repeatability is a key factor for result validity.

For those running away with the eulogies... Way too early days yet.


Gotta admit, the clocks are impressive. Every model above 3.4GHz base under full load would be very good.

Why no CPU Power or Single thread test, tho?

Currently they've only shown 8Cores+SMT favouring tests. MOAR CORES marketing.

How does that Handbrake version scale with cores/smt?

Altho bjts incredulous theory of MUCH lower power/FO4 and MUCH higher clocks than XV on 8core has failed. It's a nice sanity check.

Sent from HTC 10
(Opinions are own)

4.5GHz turbo max is still feasible. And i compared the ES situation with BD that started at 3.6 base, still possible. 4GHz base and 4.5GHz turbo is feasible only if the process is quite good and certainly not on first batches. We should wait some month at least. But 3.6 base and 4.5 turbo NOW is still feasible. i still believe on 4Ghz base as soon as process matures.

 

[KTE]

^We don't really know who has the highest performance per watt for which workload. For all intents and purposes, it could well be Zen.

Anyone assuming otherwise has Zen evidence to show, or clear bias.

Sent from HTC 10
(Opinions are own)

 

[Atari2600]

Ah. now i get it.

Erm. No. I don't think you are getting it.

If a single thread code has a number of lines of code that would repeat the order of instructions through the CPU pipe, then branch prediction can move your job on more than 1 step per clock cycle by back-loading up the right instruction rather than waiting for the instruction step to reach its conclusion.

The 4 basic steps to executing anything in a modern CPU:
1. Fetch
2. Decode
3. Execute
4. Write-back

So, if at a moment in time, your data is decoding in step 2, and you can predict based on history the required instruction, then you can initiate the fetch without waiting to step through steps 3 and 4.

Have a look at this for the level were talking about:

https://www.cs.umd.edu/class/sum2003/cmsc311/Notes/CompOrg/fetchDecode.html

 

[KTE]

It's actually not a valid point. To demonstrate the inanity of including the delta:

Zen idle: 0W
Intel idle: 106W

Zen load: 94W
Intel load: 191W

Zen delta: 94W
Intel delta: 85W

Utterly pointless.

Tested for years, esp back in Phenom days. Idle->Load AC is very inaccurate as a means to show CPU power.

Sent from HTC 10
(Opinions are own)

 

[witeken]

Zen idle: 93W
Intel idle: 106W

Zen load: 187W
Intel load: 191W


If zen is more efficient at idle, and more efficient at load, then what is your point?

If a CPU uses less power and idle, and less power at load, and has equal or better performance, then its more efficient. Simple math.

What IQ do you have or what is your education?

The power that is measured is not the Broadwell-E chip and it is not the Zen chip. It is the whole system power consumption. So concluding one chip is more efficienct than the other is a non sequitur. (If you look at my post, I did not claim one was superior than the other.)

It's like if you were to put two different SSDs in each system, say a Samsung one and a Micron one, and concluding that one is more power efficient than the other based on those 4 power numbers.

If you want to get nitty gritty about how real power measurement is -- should be -- done, you might want to read up on it before you waste any more keystrokes on nonsense like that post.

http://www.anandtech.com/show/6529/busting-the-x86-power-myth-indepth-clover-trail-power-analysis





As one can see, platform power here is more than an order of magnitude more than CPU power.


Questioning someones IQ is an insult, and thats not allowed here.
Markfw
Anandtech Moderator

 

[Dresdenboy]

About the hilarious TDP discussion.

Zen idle: 93W
Intel idle: 106W

Zen load: 187W
Intel load: 191W

Zen delta: 94W
Intel delta: 85W

So tell me, who has the highest performance per watt?

This is difficult to judge, as the Intel system only had 2 DIMMs (similar config), so it used somewhat less power due to 2 muted channels and likely a somewhat lower performance (hint, hint) caused by a possible memory bottleneck.

I think this has been analyzed before (last Blender comparison): What was the effect of using 2 vs. 4 channels in Blender?

 

[cytg111]

What IQ do you have or what is your education?..

Uncalled for.

 

[Vaporizer]

Well i am outright stunned and ryzen will succed my good old vishera cpu.

 

[realibrad]

Well, the point is that there may be more efficient platforms for the intel system. I am sure AMD spent many hours optimizing their system, and not the intel one. But in any case, the numbers are basically equal *in this particular test chosen and set up by AMD*. All the AMD fans are talking about how great the benchmarks are, etc. but basically I am only taking two things from this demo.
1. The base clocks are pretty good (better than I expected), although we dont know ultimate overclocking speeds, which will be critical in this segment.
2. Power efficiency is also competitive.

Lets face it, all the other tests, whether powerpoint slides or demos, are still staged by AMD, and just like every manufacturer, most likely reflect best case scenario for their product.

Both perfectly reasonable positions by you. The arguments by witeken are not reasonable. Questioning efficiency because of deltas is stupid.

 

[zinfamous]

What does that have to do with your statement of

You clearly answered your own question when you posted this.
Zen idle: 93W
Intel idle: 106W

Zen load: 187W
Intel load: 191W


If zen is more efficient at idle, and more efficient at load, then what is your point?

Assuming your numbers are true, Zen uses 12% less at idle, and 2% less at load.

By your delta logic, a cpu that has 200W idle, and 200W load is better than a cpu with 100W idle and 180W load because the delta for the 2nd is 80W whereas the first is 0. You do not understand numbers if you think that argument is relevant.

Also, did Intel attach a 50W light bulb to the 6900k to make the 6600k look better? I mean, the power usage of the 6900k is so high.


If a CPU uses less power and idle, and less power at load, and has equal or better performance, then its more efficient. Simple math.

One thing where I think this new paradigm of "power delta" metric might have merit, so throwing a bone to witiken here--is that this test was run with Zen clocked at 3.4 and the 6900K at 3.6? Is that right? whatever they meant by "turbo/boosting."

As awesome as it appears that Zen is--still outperforming at a lesser clock, there is an argument to be made that the Zen power draw delta from idle to load at 3.4, being much higher than the i7 draw from idle to "whatever," means that a consumer Zen hitting "whatever" is probably going to lose that 2% advantage it as at that lower clock "load." Now, if it is already performing better at a lower clock, then it is going to perform x+y better at an equal clock, but the power rating is probably going to suffer even more, no? It could be that Zen just isn't as efficient at load, but if everything as presented turns out to be true, it probably doesn't matter. That difference might be negligible (except for those desperately clinging to the last negative data point).

Zen looks to be performing beyond the Rainbows and Unicorn level of the most optimistic Red shirt-wearing fanboys, which is quite something. I would never consider myself a red-shirt, but I very much want to see AMD succeed and compete. That being said, I am thoroughly surprised with yesterday's event, as I was expecting a boring-yet-satisfactory piece of hardware that brings, at most, a new consideration to the table. At this point, I'm planning to go full AMD sometime a year from now, because I wasn't expecting to replace my Xeon E3 1231 v3 for another 1.5-2 years. Depending on what we get with a 6 core (or 8 core pricing) of Zen, I might end up spending 2x my personal historical CPU budget (right around $200 MAX if I was feeling crazy--basically, that Xeon.)

I wonder if we will start seeing real benches at CES in a few weeks? Damn, that can't come soon enough. Real Zen benches and more Vega info? could be one heck of a week for AMD if the good news keeps rolling in.

 

[superstition]

Well i am outright stunned and ryzen will succed my good old vishera cpu.

I think the benchmark is questionable in relation to Piledriver — in terms of the big picture. A 4 core 4 thread ancient Lynnfield i5 should not be beating or matching a 4–4.6 GHz FX except in tests that strongly favor single-threaded performance and/or FPU performance.

Lynnfield does have better single-thread performance than Piledriver but certainly not equivalent multithreaded performance when compared with Piledriver at 4.6 GHz! Even Cinebench R15, which some have said is designed to favor Intel, shows this. (However, it could be that the lack of AVX in the Lynnfield is an issue with CB R15.)

And yet, at 3.8 I got the same 200 sample score as someone with a 4.6 GHz FX and my 100 sample score tied someone else's 4 GHz stock 8350.

 

[realibrad]

What IQ do you have or what is your education?

The power that is measured is not the Broadwell-E chip and it is not the Zen chip. It is the whole system power consumption. So concluding one chip is more efficienct than the other is a non sequitur. (If you look at my post, I did not claim one was superior than the other.)

It's like if you were to put two different SSDs in each system, say a Samsung one and a Micron one, and concluding that one is more power efficient than the other based on those 4 power numbers.

If you want to get nitty gritty about how real power measurement is -- should be -- done, you might want to read up on it before you waste any more keystrokes on nonsense like that post.

http://www.anandtech.com/show/6529/busting-the-x86-power-myth-indepth-clover-trail-power-analysis





As one can see, platform power here is more than an order of magnitude more than CPU power.

You seem to be getting upset because you are questioned about your logic.

We do know that the AMD system in this context used slightly less power. There could be better system setups that use less, but this could also be perfectly reasonable too. You dont know.

What I called into question is your use of delta measurements as anything meaningful. The biggest gap in the delta is the idle usage. You for some reason highlighted the delta in idle vs load as if that somehow means something in this argument and it does not. Your argument is now that AMD put in different hardware beyond the CPU to make the Intel system look worse. While that is possible, you have no reason to believe this. You are now going into the realm of conspiracy.

So, why do you believe delta measurements are important to this argument? Its something to look at for other reasons, but not perf/watt.

 

[sirmo]

Also, for those interested in power consumption, during the Blender demo the two systems consumed roughly the same amount of power, with Zen having the edge. 100Mhz more for Zen and the two systems would have been evenly matched as far as power is concerned.

Where do you get 100Mhz more for Zen? Just curious. Stock 6900k is 3.2Ghz base so at base it would be 200Mhz, which is 6.25% clock advantage. However a lot of reviews I read about 6900k says that under full load all 8 cores still turbo to 3.5Ghz.. Which would mean 6900k had a 3% clock advantage.

Would be cool to get a confirmation from one of 6900k owners.

 

[Gikaseixas]

What IQ do you have or what is your education?

Wow... was this really necessary?

 

[Gikaseixas]

DP

 

[bjt2]

Erm. No. I don't think you are getting it.

If a single thread code has a number of lines of code that would repeat the order of instructions through the CPU pipe, then branch prediction can move your job on more than 1 step per clock cycle by back-loading up the right instruction rather than waiting for the instruction step to reach its conclusion.

The 4 basic steps to executing anything in a modern CPU:
1. Fetch
2. Decode
3. Execute
4. Write-back

So, if at a moment in time, your data is decoding in step 2, and you can predict based on history the required instruction, then you can initiate the fetch without waiting to step through steps 3 and 4.

Have a look at this for the level were talking about:

https://www.cs.umd.edu/class/sum2003/cmsc311/Notes/CompOrg/fetchDecode.html

I know how OOO and superscalar architecture work. What i was saying is than the same situation apply to INTEL.

 

[realibrad]

One thing where I think this new paradigm of "power delta" metric might have merit, so throwing a bone to witiken here--is that this test was run with Zen clocked at 3.4 and the 6900K at 3.6? Is that right? whatever they meant by "turbo/boosting."

As awesome as it appears that Zen is--still outperforming at a lesser clock, there is an argument to be made that the Zen power draw delta from idle to load at 3.4, being much higher than the i7 draw from idle to "whatever," means that a consumer Zen hitting "whatever" is probably going to lose that 2% advantage it as at that lower clock "load." Now, if it is already performing better at a lower clock, then it is going to perform x+y better at an equal clock, but the power rating is probably going to suffer even more, no? It could be that Zen just isn't as efficient at load, but if everything as presented turns out to be true, it probably doesn't matter. That difference might be negligible (except for those desperately clinging to the last negative data point).

Zen looks to be performing beyond the Rainbows and Unicorn level of the most optimistic Red shirt-wearing fanboys, which is quite something. I would never consider myself a red-shirt, but I very much want to see AMD succeed and compete. That being said, I am thoroughly surprised with yesterday's event, as I was expecting a boring-yet-satisfactory piece of hardware that brings, at most, a new consideration to the table. At this point, I'm planning to go full AMD sometime a year from now, because I wasn't expecting to replace my Xeon E3 1231 v3 for another 1.5-2 years. Depending on what we get with a 6 core (or 8 core pricing) of Zen, I might end up spending 2x my personal historical CPU budget (right around $200 MAX if I was feeling crazy--basically, that Xeon.)

I wonder if we will start seeing real benches at CES in a few weeks? Damn, that can't come soon enough. Real Zen benches and more Vega info? could be one heck of a week for AMD if the good news keeps rolling in.

Quite reasonable. One thing I would say is that the idle usage may be due to other things than frequency scaling. If there are power saving systems that kick in to turn things off, then increasing the frequency may not follow the linear usage from idle to load. What we would need from there is to see power usage while in boost that we simply dont have. The fact that under load the CPU uses around the same power is a sign that they probably wont be far off in scaling for OC. These are complex though, so we wont know until we get one.

What witiken did was to question the perf/watt because of the delta difference. That is absurd and shows a lack of understanding. Everything else is a shift away from that point.

 

[USER8000]

Uncalled for.

This.

 

[zinfamous]

What IQ do you have or what is your education?

Look, man: it seems that you lost this argument long ago--why make it worse?

 

[cytg111]

Its easy to manipulate an experiment for a desired result.

Hence why independent repeatability is a key factor for result validity.

For those running away with the eulogies... Way too early days yet.


Gotta admit, the clocks are impressive. Every model above 3.4GHz base under full load would be very good.

Why no CPU Power or Single thread test, tho?

Currently they've only shown 8Cores+SMT favouring tests. MOAR CORES marketing.

How does that Handbrake version scale with cores/smt?

Altho bjts incredulous theory of MUCH lower power/FO4 and MUCH higher clocks than XV on 8core has failed. It's a nice sanity check.

Sent from HTC 10
(Opinions are own)

Yea... no single threaded benches yet... a little weird.

 

[Glo.]

I am wondering, why everyone is talking about the Blender benchmark, whereas AMD demoed one more bench: Handbrake.

And that is also possible to test on your own, and compare notes.

 

[superstition]

I am wondering, why everyone is talking about the Blender benchmark, whereas AMD demoed one more bench: Handbrake.

And that is also possible to test on your own, and compare notes.

What file do we use to replicate their test?

 

[coercitiv]

Where do you get 100Mhz more for Zen?

I only said a minor speed bump for Zen would put the two systems in that demo at equal power usage levels. I did not say anything about the 6900K clocks.