AMD Zen “RYZEN” CPUs Detailed – 8 Cores, 3.4Ghz+ & Auto Overclocking With “XFR”

Page 12 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Soulkeeper

Diamond Member
Nov 23, 2001
6,735
155
106
It's good question, does SMT in Zen have higher ultilization?

IF this is the case, Zen's power consumption might be much higher when SMT is on, thus it would have potencially less efficiency compared to Intel's. For example, BD/PD/SR has significantly higher consumption when go from 2C to 2M4C.

But AMD's own demo tell us, Zen's power consumption is not high. This COULD imply a low utilization of AMD's SMT.

Just a thought, I could be wrong.:rolleyes:

If it has higher utilization, then that could just mean faster execution ie: race to idle
 

cytg111

Lifer
Mar 17, 2008
25,996
15,443
136
In some scenarios it is that simple, and it could be expressed in a simple formula. But since Zen's SMT yield in Blender lies somewhere between -100% and +100%, there is a wide field for speculation about ST performance. ;)

BTW with not too many memory references (AGU+L/S contention), FP code can see nice SMT yields AND good ST performance for the same code. Explanation: FP ops have multi cycle latencies. If there are only one or few dependency chains in the code, these latencies offer enough opportunities for another FP thread to slip in. Another source of opportunities are cache misses, causing long stalls. And power constraints in MT mode may hide the true SMT yield. Example with CB R15:
http://www.pcworld.com/article/3039...es-you-really-need-for-directx-12-gaming.html
2C -> 2C HT: SMT yield 37%
8C -> 8C HT: SMT yield 32%

Does 37% yield indicate bad ST performance for Haswell?

Even SB with all cores had some nice SMT yields for FP heavy stuff. Sometimes also just 0% -> no correlation with ST performance.

Yes I get all that (well maybe not *all*) yet in a world governed by laws like thermo dynamics conservation of energy (conversation of performance) and the fact that we know that Broadwell equals Zen for the MT demo... we could put it like this

P Performance
b Broadwell
z Zen
zc Zen core
zt Zen hyperthread
bc Broadwell core
bt Broadwell hyperthread

P(b) = P(z)
and
P(b) = (bc + bt)
and
P(z) = (zc + zt)
then
bc + bt = zc + zt
then
P(zc) = bc + bt - zt

Best case scenario zt = 0.
 

unseenmorbidity

Golden Member
Nov 27, 2016
1,395
967
96
It does come with some ram limits, though. It seems odd to limit what looks like a very good chip.

This isn't an HEDT CPU, and it's not the market they are going for. If you want 64 gb of memory for video editing or something, then you should go with intel.

And next time they might show us 7zip or fritz benchmark and we will end up thinking that zen will have 20% more IPC then sky/kaby...
I'm not talking about cherry picking as such,but avoiding any ST test should be a red flag for anyone.

If they really are that close in ST then boosts and such would be irrelevant they would show a test, without numbers just like the blender test, just to brag about it.

They can't, because they are still fine tuning the boost frequency and/or technology.
 
Last edited:

coercitiv

Diamond Member
Jan 24, 2014
7,345
17,391
136
It's good question, does SMT in Zen have higher utilization?

IF this is the case, Zen's power consumption might be much higher when SMT is on, thus it would have potentially less efficiency compared to Intel's.
In what distorted alternate reality does better SMT utilization lead to less efficiency?! If that were true SMT would be counterproductive from the start, leading to less efficiency across the board.

For MT loads SMT offers ~ linear increase of performance & power usage, while frequency drop available in order to normalize performance to pre SMT increase levels will result in exponential power usage decrease.

Just imagine 10% more performance for 10% more power, allowing you to drop MT frequency by 9%!
 

deasd

Senior member
Dec 31, 2013
603
1,033
136
In what distorted alternate reality does better SMT utilization lead to less efficiency?! If that were true SMT would be counterproductive from the start, leading to less efficiency across the board.

For MT loads SMT offers ~ linear increase of performance & power usage, while frequency drop available in order to normalize performance to pre SMT increase levels will result in exponential power usage decrease.

Just imagine 10% more performance for 10% more power, allowing you to drop MT frequency by 9%!

You've misunderstood.

More utilization = more resources to be used = more heat
Less utilization = less resources to be used = less heat

I just replied to the doubt that Zen's SMT could bring more % performance than Intel's, implying Zen's single threading is weaker, but from Zen's power consumption we know that Zen's SMT MIGHT not utilize that much resources.

What you said is correct, SMT could bring higher efficiency than multi-core, but it doesn't related to what I said.
 

TheELF

Diamond Member
Dec 22, 2012
4,027
753
126
Another source of opportunities are cache misses, causing long stalls. And power constraints in MT mode may hide the true SMT yield. Example with CB R15:
http://www.pcworld.com/article/3039...es-you-really-need-for-directx-12-gaming.html
2C -> 2C HT: SMT yield 37%
8C -> 8C HT: SMT yield 32%

Does 37% yield indicate bad ST performance for Haswell?
We all know that no workload is parallel only they always have at least a small amount of serial code to keep an eye out over all the other threads and to sort all the data that is shared to the threads and comes back from them.
The more threads there are the more work this one serial/sequential thread has to do the less efficient HT/SMT becomes.
jaiKV.png



Even SB with all cores had some nice SMT yields for FP heavy stuff. Sometimes also just 0% -> no correlation with ST performance.
Nobody is arguing that HT slows down the execution of a non HT thread...

Think of it in simple terms if 100% is one core/thread on hass/broadwell and in some program (blender for example) you get 100% + 30% from HT then one core basically runs at 130% of the speed it's capable of(related to this program) .

Now if one core/thread on ZEN is only 65% of one hass/broadwell core but their SMT implementation runs a second thread on the same core at another 65% it adds up to the same 130%

So the question is if zen actually has the same IPC as has/broadwell or if it is well below but with much better throughput.
Look at the presentation again 25% of the 40% uplift is the new technologies they introduced, smart prefetch and neural whatever_it_is, so they outright telling us that IPC as we discuss it(hard cruel real world IPC) is only 30% over excavator, 40% if you count everything.
With one thread/core at 30-40% over excavator you can't reach broadwell levels,no where near actually,but with two threads/core it adds up to 60-70% IPC gain which would more or less be on the level of IPC improvement people are dreaming of in this (and all other) thread(s).
So it all comes down to what form of SMT they use partitioned as in the textbook passage(close to 100% SMT scaling) I showed or unpartitioned like HTT where ever thread just grabs what it can ( ~30% in most bench cases but also close to 100% in most games that scale well with cores hence why the i3 is still good for games) .

And yes since lucy lu said that they put senseMI towards IPC they are not talking about the same IPC we are talking about.
Nobody cares that the skylake i3 has way better min fps in bf1 then the haswell i3 after adjusting for clock speeds all people care about is the benches showing only ~5% IPC improvements between haswell and skylake.
 
  • Like
Reactions: cytg111

coercitiv

Diamond Member
Jan 24, 2014
7,345
17,391
136
More utilization = more resources to be used = more heat
more utilization = more resources to be used = more heat = more performance = better efficiency for normalized performance

I just replied to the doubt that Zen's SMT could bring more % performance than Intel's, implying Zen's single threading is weaker
Next in line we'll discuss how the superior GloFo 14nm node is helping AMD offset the efficiency disadvantage Zen has against Broadwell. (oh boy, if I see this idea promoted on the forums I'll laugh till I poop myself)

It wasn't long ago that the unanimous opinion among certain fans was that AMD would have less SMT utilization, not higher. In fact every metric or information we had so far has been anally scrutinized and turned upside down in order to minimize estimated Zen performance.

Zen was supposed to:
  • be a low power product. Main argument was AMD could not afford design of a high performance product. Yes, right here on the forums.
  • have Sandy Bridge like ST performance. Main argument was AMD could not afford R&D expense & talent to design a top notch high performance product.
  • stay under or at max 3Ghz for base clocks. Main argument was GloFo 14nm would fail, like it always does.
  • have inferior SMT scaling to Intel's. Main argument was this is the first time AMD implements SMT, Intel has far more experience and iterations.
Every one of these highly debated subjects has been debunked. At this point we know AMD has a decent product on their hand, one which defies every prediction based on their previous ability to execute. This is a time for leaks and benchmarks, it's no longer a time for wishful thinking.
 

cytg111

Lifer
Mar 17, 2008
25,996
15,443
136
...Every one of these highly debated subjects has been debunked. At this point we know AMD has a decent product on their hand, one which defies every prediction based on their previous ability to execute. This is a time for leaks and benchmarks, it's no longer a time for wishful thinking.

I think you are getting ahead of the evidence here, you may turn out to be right, but as of right now I dont think the evidence *quite* supports all of it. It is looking good, yes it is, but I still doubt that it is GloFo and i doubt singlethreaded perf stacks up as well as multi .. I would love to cross those two off as well, but cannot do that without the evidence, no sir.
 

Abwx

Lifer
Apr 2, 2011
11,879
4,864
136
So it all comes down to what form of SMT they use partitioned as in the textbook passage(close to 100% SMT scaling) I showed or unpartitioned like HTT where ever thread just grabs what it can

More non sense..

100% SMT scaling means that it would have 10% lower ST perf (lol) than XV since this latter is slightly faster than a 2/4T HW in a single pic rendering using Blender...
 

inf64

Diamond Member
Mar 11, 2011
3,884
4,692
136
"Close to 100% SMT scaling" -> just LOL.

Some posters here are getting desperate :/
 

cytg111

Lifer
Mar 17, 2008
25,996
15,443
136
He is not arguing that Zen has 100% SMT scaling, he is drawing one extreme vs. another and implies that the truth is somewhere in between. This is a valid scientific method, examine the extremes to get a grasp of what is possible and whats not.
 
  • Like
Reactions: DarthKyrie

LTC8K6

Lifer
Mar 10, 2004
28,520
1,576
126
This isn't an HEDT CPU, and it's not the market they are going for. If you want 64 gb of memory for video editing or something, then you should go with intel.

So that's why all the comparisons to BW-E then? Because Ryzen is not HEDT?

If Ryzen is not HEDT, then it's the biggest CPU smash hit of all time, and I can't wait to see the HEDT version.
 
  • Like
Reactions: Phynaz

Abwx

Lifer
Apr 2, 2011
11,879
4,864
136
He is not arguing that Zen has 100% SMT scaling, he is drawing one extreme vs. another and implies that the truth is somewhere in between. This is a valid scientific method, examine the extremes to get a grasp of what is possible and whats not.

That s not valid at all, a Zen core in SMT has about the same throughput as an XV module, so he s saying that it s quite possible that Zen has same ST IPC than a XV core, and even lower since a module has 10% penalty respectively to a XV single core.

From the start that s total non sense, but well, if you call this a valid methodology to start with mistakes all over the place, why not..?..

"Close to 100% SMT scaling" -> just LOL.

Some posters here are getting desperate :/

AMD said that they exeeded 40%, so if IPC was increased by 45-50% then it means that their SMT bring 50-45%, at least in Blender.
 
Last edited:

cytg111

Lifer
Mar 17, 2008
25,996
15,443
136
so he s saying that it s quite possible that Zen has same ST IPC than a XV core, and even lower since a module has 10% penalty respectively to a XV single core.

Yes. It is possible. Careful not to confuse possible with probable. Zens ST performance is probably very very good while still possible that is not quite what we hoped for.
If TheELF ment to say that Zens ST performance will most defn. suck major .. then i have misread the situation. Dont think I have.
 

deasd

Senior member
Dec 31, 2013
603
1,033
136
With one thread/core at 30-40% over excavator you can't reach broadwell levels,no where near actually,

I won't emphasis it again: 40% more IPC is not equal to 40% more performance.

For example, going from AVX128 to AVX256 you get 100% theoretical performance uplift, but in practical situation you always cannot achieve this goal, sometimes maybe just 30%-50% uplift, in this situation IPC would be lower than legacy mode due to the horrible latency of longer instruction.

I don't know how Blender works, but looks like AMD seriously optimizing latency of long instruction window in Zen. OTOH it makes no sense to compare IPC between Zen and EXV in some heavy workload like this.

but with two threads/core it adds up to 60-70% IPC gain which would more or less be on the level of IPC improvement people are dreaming of in this (and all other) thread(s).

And this is what you made up. IPC is single threaded metric.
 
Last edited:

Abwx

Lifer
Apr 2, 2011
11,879
4,864
136
Yes. It is possible. Careful not to confuse possible with probable. Zens ST performance is probably very very good while still possible that is not quite what we hoped for.
If TheELF ment to say that Zens ST performance will most defn. suck major .. then i have misread the situation. Dont think I have.


It is impossible, from the start since they stated 40% for a single thread, i dont understand why you keep supporting such irrelevancies that bring nothing to the debate, quite the contrary..
 

cytg111

Lifer
Mar 17, 2008
25,996
15,443
136
It is impossible, from the start since they stated 40% for a single thread, i dont understand why you keep supporting such irrelevancies that bring nothing to the debate, quite the contrary..
Please dont question my relevance to this thread, I dont question yours and please dont drag it down to that level, I wont follow and ultimately ill hit the report button.

You quote 40% as proof of something, those 40% is directly from AMD's marketing slides, that is not proof of anything at all. Besides, if those 40% stands, it supports TheELF's argument even more.
 

Abwx

Lifer
Apr 2, 2011
11,879
4,864
136
Please dont question my relevance to this thread, I dont question yours and please dont drag it down to that level, I wont follow and ultimately ill hit the report button.
.
I m questionnng the relevancy of the made up theories you re saying are valid, KTE s theory is wrong from the start because it would imply that Zen has same IPC as excavator, keeping supporting such non sense automatically make someone irrelevant to the technical discussion.

Or rare you saying that i can state that IPC increased by 200% and SMT gain is -50%, that s possible according to your standard of possibilities, but would it be really relevant..?..


.

You quote 40% as proof of something, those 40% is directly from AMD's marketing slides, that is not proof of anything at all. Besides, if those 40% stands, it supports TheELF's argument even more.

I dont see the point of using ignorance as an argument, AMD gave a lot of infos and we are talking about thoses infos while you are saying that it s bs and that we should use non sense as basis for the discussion, that is dubbious theories, that are contradicted by AMD s own statements..

Besides if thoses 40% hold that invalidate completely his "theory" of possibly 100% SMT scaling, that would make 2.8x an XV core throughput, so you re using an agument that is a demonstration that you are not even undestanding your own words, but i guess that it s not about being accurate..
 
Last edited:
  • Like
Reactions: DarthKyrie

coercitiv

Diamond Member
Jan 24, 2014
7,345
17,391
136
I think you are getting ahead of the evidence here, you may turn out to be right, but as of right now I dont think the evidence *quite* supports all of it.
Pick a bullet in that list that is not yet debunked.

It is looking good, yes it is, but I still doubt that it is GloFo
Ofc it's not GloFo, that was a sarcastic outburst meant to emphasize the kind of explanation some might invent to address the possibility of Zen being a good product. And when I say good/decent, I don't mean better than 6900K, but rather competitive (fast enough / cool enough / cheap enough)

and i doubt singlethreaded perf stacks up as well as multi..
I never stated Zen has better ST perf/clock than BW, and I doubt it does, since the very little gaming data we have showed 6900K performing better (and that only hints at ST perf, and only for a specific type of load). But don't you see the irony of having to invoke better SMT implementation in order to explain equal throughput at equal power levels?

AMD shouldn't be even remotely this close. They should be bringing guns to a laser fight.
 

cytg111

Lifer
Mar 17, 2008
25,996
15,443
136
Pick a bullet in that list that is not yet debunked.

Ofc it's not GloFo, that was a sarcastic outburst meant to emphasize the kind of explanation some might invent to address the possibility of Zen being a good product.

Is your list representative for the state of Zen one way or the other? Past failed prophesies doesnt warrant the correctnes of current ones.
Only argument I am making is that until you actually open the the box and look at that damned thing that cat is both dead and alive.
We simply wont know until we see ST performance
We simply wont know until we see final clocks
We simply wont know until we see production ramping up
 

krumme

Diamond Member
Oct 9, 2009
5,956
1,596
136
Pick a bullet in that list that is not yet debunked. But don't you see the irony of having to invoke better SMT implementation in order to explain equal throughput at equal power levels?

.

I guess most can see the irony.
As that in no one on this planet beforehand would say amd smt would beat Intel. Suddenly its a possibility. Lol.
I can't simply grasp that possibility even with 4x fp units.
Yet on the same way its impossible to grasp st perf to be on bwe level. As shown by the high base we were shown.
It crazy stuff anyway its viewed.
 
Last edited:

krumme

Diamond Member
Oct 9, 2009
5,956
1,596
136
Go back and read the stuff page 1-3 from sweepr zen bm thread. Like

"Does the Jim Keller pixie dust come with any sort of a guarantee?"

I think that was spot on back then. Now the answer is just:
Yes.
Lol.
It simply looks like pixie dust to me. Perhaps i have even gone a bit religious.
Nice to see most having a good time enjoying it. Its what enthusiasm is.
 
  • Like
Reactions: DarthKyrie

coercitiv

Diamond Member
Jan 24, 2014
7,345
17,391
136
Is your list representative for the state of Zen one way or the other? Past failed prophesies doesnt warrant the correctnes of current ones.
Only argument I am making is that until you actually open the the box and look at that damned thing that cat is both dead and alive.
You do realize you're contradicting with me by saying the exact same thing, right?

That list was the basis of my last argument: all bets are off, from now on it's leaks & benches that matter.