AMD FX "Bulldozer" Review - (4) !exclusive! Excuse for 1-Threaded Perf.

[BeepBeep2]

Windows 7 is already handling things like this for Intel processors with HT, using real cores first and logical cores later.

However, according to AMD there are situations where you don't even want this behavior.
Take a look at the first two pictures at THG:
http://www.tomshardware.co.uk/fx-815...-32295-23.html

Because of the shared L1-Cache it makes indeed sense that in some cases it can be faster to use the whole module instead of splitting things up and utilize two modules partially. This means that the scheduler has to be more intelligent though, as it's not enough to just assign each new task to a new core like now, instead it must be able to guess which tasks should be grouped to one module and which should be split over two (more) modules.

I'm no coder but I can imagine easier projects than making the scheduler aware of such a complex problem.

There is no real or logical core in BD.
There are clusters, simple as that.
When you disable a cluster in BIOS, you do the same thing as AMD's diagram.

AMD's diagram
Core 0 - shared
Core 1 - one cluster
Core 2 - shared
Core 3 - one cluster (uses all resources for 1 thread)

What were doing
Core 0 - one cluster
Core 1 - disabled
Core 2 - one cluster
Core 3 - disabled

[pumero]

I'm aware of that and I never said that it works like that on BD.
At the moment Windows sees the processor as having 8 real cores and assigns the tasks accordingly but doesn't care (know) about the whole module thing.

[Spectrobozo]

Jack, hardware.fr already tried something along those lines with 4m/4t and 2m/4t ,both with Turbo on. In 1st case maximum turbo for all 4 "threads" is 3.9Ghz since all modules are running. In second case it's 4.2Ghz across 2 modules(4 threads). The % difference in Turbo clock(~7%) is not nearly enough to make up for sharing losses as can be seen here:
http://www.hardware.fr/articles/842-...windows-8.html
Attachment 121226

4m/4t is 26% faster(!) than 4m/2t at fixed 3.6Ghz and 15% faster when both are running their maximum Turbo modes allowed. Now comes the power draw story.
If you look at the power draw you will see the faster config is 20% more power hungry and I suspect this is the reason why AMD didn't configure the core priorities in that way. I think when PD arrives,power draw will go down sufficiently in order to schedule the threads the faster way and still get good power numbers. Still,with present BD core, for 20% more power you gain 26% more performance this way,not a bad tradeoff. If GloFo would get their act together and make possible for AMD to produce 3.6Ghz 5 module PD core with this thread affinity capability,this thing could very well be significantly more powerful than Thuban ,even in ST at fixed clock and noticeably more powerful than BD in both ST and MT with both Turbo on and off.

By the way,great thread DGLee

interesting,

so on the graphic 4t and 8t are at the default CPU settings (4m/8c), only with the software set to use 4 or 8 threads?

the increase from the default settings with turbo off, to the 4m/4c was quite good, 13%, with turbo on more like 7-8%?


there is a lot going on to determine which is the optimal performance/power usage in how you handle the modules, threads, turbo...

[Hornet331]

Jack, hardware.fr already tried something along those lines with 4m/4t and 2m/4t ,both with Turbo on. In 1st case maximum turbo for all 4 "threads" is 3.9Ghz since all modules are running. In second case it's 4.2Ghz across 2 modules(4 threads). The % difference in Turbo clock(~7%) is not nearly enough to make up for sharing losses as can be seen here:
http://www.hardware.fr/articles/842-...windows-8.html
Attachment 121226

4m/4t is 26% faster(!) than 4m/2t at fixed 3.6Ghz and 15% faster when both are running their maximum Turbo modes allowed. Now comes the power draw story.
If you look at the power draw you will see the faster config is 20% more power hungry and I suspect this is the reason why AMD didn't configure the core priorities in that way. I think when PD arrives,power draw will go down sufficiently in order to schedule the threads the faster way and still get good power numbers. Still,with present BD core, for 20% more power you gain 26% more performance this way,not a bad tradeoff. If GloFo would get their act together and make possible for AMD to produce 3.6Ghz 5 module PD core with this thread affinity capability,this thing could very well be significantly more powerful than Thuban ,even in ST at fixed clock and noticeably more powerful than BD in both ST and MT with both Turbo on and off.

By the way,great thread DGLee

Yes 4C/4M is 26% faster then 4C/2M, but i guess jack meant what happens when you compare fixed threads (aka 4C/4M) with OS scheduling. ANd when you I look at your posted graph its actually not that bad. I assume that 4T means 4 threads on the fully enabled 8 core cpu.

So that means you lose 7-13% to the ideal condition where each module runs only one thread. Not that of a difference, it its surprisingly close to the numbers AMD posted with the improved win8 scheduler.

Plus that doesn't show how well one Moduel compares to one P2 core.

[informal]

P2 scales almost perfectly to 2 threads in most workloads. Bulldozer doesn't.My point is that many workloads in benchmark suits are lightly threaded and this 4m/4t,if it could be switche on by a patch(or special bios hook via AOD for instance) ,could bring the average number Zambezi gets by a rough 10-12%. This would easily put Zambezi over Thuban in almost all workloads(1-8 thread,Turbo enabled of course) since now you would have 15-20% better (then before) results in 1-4 thread workload cases.

As for single core performance,maybe the OP can run some tests with affinity set to only one core via task manager. We can then compare the results with Phenom II.

[Hornet331]

You only get way with that if you get higher clocks and even then its hard... a 3.3ghz phenom beats a 3.9ghz BD in singlethreaded apps... and only pulls slightly ahead in MT apps due to the better Turbo implementation.

Just look at the 2 pages of the anand tech review:
http://www.anandtech.com/show/4955/t...x8150-tested/4
http://www.anandtech.com/show/4955/t...x8150-tested/7

If they can't increase the ST performance above P2 or provide the clockspeed to outperform P2 in ST and maintain this clockspeed in 1-6 threads, they gona loose to P2.

Sure it depends on apps, but if we go after the CB R11.5 results BD needs at least to hold 3.9-4.0 ghz constantly till 6 threads to be faster then a P2. it doesnt matter if you hook the thread to specefic moduls, when one module is already slower then one core of P2 you cant expect that 2 slower moduls will beat 2 faster cores...

[informal]

Cinebench is only one test... Why are you using it as end all be all scenario?

[Hornet331]

No read what I wrote... I said it depends on the apps. There are apps where BD is very close or even faster compared to P2 in ST and wins there due to clock speed or core advatnage and naturally also wins there in MT. The problem are the situations where ST performance just lags to much behind which can't be fixed by the clock speed (larger then 18%) and it gets painfully obviouse if the workload is only "slightly" threaded (less then 4 or 6 threads, depending to with P2 you compare it).

There are hardly any reviews out there that for e.g. use lame or itunes in there test suite anymore, but there are still million of people out there that use this apps (lame is 2T afair itunes is even only ST)

[informal]

Yeah IPC needs to be improved,that is for sure. I hope they hit the 10% listed for Piledriver from the FX Next slide. There are definitely some bottlenecks ,especially for legacy(old) code, primarily in the front end and L1 cache. If they manage to fix some of these issues and if GloFo improves their 32nm process,we would welcome Piledriver with more favorable reviews.

[PatRaceTin]

fun core disable

[awdrifter]

Good work. Let's hope AMD will be able to correct this.

[demonkevy666]

http://vr-zone.com/articles/amd-fx-8...-/13694-6.html

the advance shows something.

is sandy bridge really 25-30% faster then nehalem?

i seen a I7 875K scores 1.21 in single thread with turbo at 3.6ghz vs 3.8ghz of the 2600k which scores 1.52.

[The Stilt]

I did some testing with the "native" 4CU/4C configuration a while ago, and saw significant performance increase over 2CU/4C configuration.
When the compute units are set to single core mode, the bios versions available back then did not change the per core L3 partition/allocation size accordingly (from 1MB to 2MB). I will retest this as soon as the new bios I´ve requested is available.

[BeepBeep2]

Yeah IPC needs to be improved,that is for sure. I hope they hit the 10% listed for Piledriver from the FX Next slide. There are definitely some bottlenecks ,especially for legacy(old) code, primarily in the front end and L1 cache. If they manage to fix some of these issues and if GloFo improves their 32nm process,we would welcome Piledriver with more favorable reviews.

I am wondering if that 10% increase comes from a 1.25% increase in performance over each thread (The way AMD has been marketing, I wouldn't doubt it at all)

AMD needs a 45-50% gain to compete with intel in single thread, at the same clock scaling and unless Intel designs Netburst 2 it will only get worse.
The IPC vs core speed gap went from 25% and 15% clocks to 50% and 5% clocks...not what I was expecting at least.

I spoke so long about how competitive STARS is from a performance per mm2 perspective vs Nehalem...and to be honest, it is also very competitive on a performance per watt perspective as well. They should have shrunk X6, adopted the tweaked and refined Llano "STARS" core with it's 5% IPC boost, doubled the L2 cache (to give 1MB per core), doubled L3 to 12MB...refined memory controller like they did with BD...increased core speed to 3.7/3.9 Turbo/4.2 ST Turbo or so.

That would be a competitive CPU, would it not?

[demonkevy666]

I am wondering if that 10% increase comes from a 1.25% increase in performance over each thread (The way AMD has been marketing, I wouldn't doubt it at all)

AMD needs a 45-50% gain to compete with intel in single thread, at the same clock scaling and unless Intel designs Netburst 2 it will only get worse.
The IPC vs core speed gap went from 25% and 15% clocks to 50% and 5% clocks...not what I was expecting at least.

I spoke so long about how competitive STARS is from a performance per mm2 perspective vs Nehalem...and to be honest, it is also very competitive on a performance per watt perspective as well. They should have shrunk X6, adopted the tweaked and refined Llano "STARS" core with it's 5% IPC boost, doubled the L2 cache (to give 1MB per core), doubled L3 to 12MB...refined memory controller like they did with BD...increased core speed to 3.7/3.9 Turbo/4.2 ST Turbo or so.

That would be a competitive CPU, would it not?

you don't need the L3 cache then at all. it only usually adds 2-5% increase.

[BeepBeep2]

you don't need the L3 cache then at all. it only usually adds 2-5% increase.

5% IPC from refined core, 2-5% from L3, 2-5% from L2...
Percentages start adding up after a while...best case scenario that is 15% IPC...

If they could have gotten 6 STARS cores to do 4.6 Ghz OC'ed (so 400 Mhz increase) while gaining ~10% IPC through Llano's refined core + 2x L2 and L3...

I wouldn't doubt for a second it would be praised.

[zoson]

wrong thread. sorry.

[The PyroPath]

Would it be possible to test an intermediate situation. 6C/6T or if you prefer, 6C/6M. So enable all clusters in 2 modules, enable half the clusters in the 2 other modules? Would be nice to see what the impact of this is so you can compare between the 4C/2M, 4C/4M, etc options and see hoe that scales...

[The Stilt]

Would it be possible to test an intermediate situation. 6C/6T or if you prefer, 6C/6M. So enable all clusters in 2 modules, enable half the clusters in the 2 other modules? Would be nice to see what the impact of this is so you can compare between the 4C/2M, 4C/4M, etc options and see hoe that scales...

Yes..
The only trouble is that there is physically only four compute units ("modules") in a Zambezi node.
We would need to have a Valencia for this test.

[Andi64]

I am wondering if that 10% increase comes from a 1.25% increase in performance over each thread (The way AMD has been marketing, I wouldn't doubt it at all)

AMD needs a 45-50% gain to compete with intel in single thread, at the same clock scaling and unless Intel designs Netburst 2 it will only get worse.
The IPC vs core speed gap went from 25% and 15% clocks to 50% and 5% clocks...not what I was expecting at least.

I spoke so long about how competitive STARS is from a performance per mm2 perspective vs Nehalem...and to be honest, it is also very competitive on a performance per watt perspective as well. They should have shrunk X6, adopted the tweaked and refined Llano "STARS" core with it's 5% IPC boost, doubled the L2 cache (to give 1MB per core), doubled L3 to 12MB...refined memory controller like they did with BD...increased core speed to 3.7/3.9 Turbo/4.2 ST Turbo or so.

That would be a competitive CPU, would it not?

Completely agree. I said they should have come with something like "Phenom II X8", but you are right, even a Stars X6 at 32nm HKMG, 3.6/4.2Ghz with improved Memory Controller would be nicer than this, and I would buy it.

In this scenario I'm waiting for i7-3820 to decide between that and 2600K...

[EvilBlitz]

How about with the 4 modules, can you disable half the clusters in 2 of the modules. Then could you make the two modules with half the cluster disabled the first 2 primary cores in a ST and lightly MT apps.


Boost ST and lightly threaded apps while maintaining good heavy threaded performance maybe?


The Stilt, in this combo could it change the cache allocated on the fly so if only 2 threads it could then double it?

[The PyroPath]

How about with the 4 modules, can you disable half the clusters in 2 of the modules. Then could you make the two modules with half the cluster disabled the first 2 primary cores in a ST and lightly MT apps.


Boost ST and lightly threaded apps while maintaining good heavy threaded performance maybe?


The Stilt, in this combo could it change the cache allocated on the fly so if only 2 threads it could then double it?

That was kind of what I meant...

[charged3800z24]

With AOD, you can set applications that are single threaded to only run on a specific core using profiles and run a series of test once set and see how it does using said Core configurations in bios

[dess]

Windows 7 is already handling things like this for Intel processors with HT, using real cores first and logical cores later.

That's fine (except your wording is inaccurate). Somehow we should trick it to use this method for BD, as well...

However, according to AMD there are situations where you don't even want this behavior.

It depends on if the penalty of forcing those closely related threads to communicate through L3 (instead of L2) is more or less than the gain on the lack of sharing resources. It seems most applications only benefits from it:

img0033832.gif

So, there could be a little patch that simply enables scheduling a' la SMT in Win7, that it already supports (if true)...

Quoted from the article:

According to AMD, Windows 8 will more intelligently align threads so that, when they can benefit from sharing a module, they will. The implication is that when two threads can be consolidated onto one module (despite the fact that they’re forced to share resources), putting an entire module to sleep and potentially enabling a higher p-state (a faster Turbo Core setting) outweighs any performance penalty tied to sharing.

And so the default behaviour will be separation (contrary to what JF said all along)? Would be just stupid if not... Of course, power consumption is higher because more modules are active, but here we can see also that with turbo enabled the the energy efficiency is really the same...

Well, unless there is a fix coming (HW or SW or both) that largely improves on the penalty of sharing resoruces. Just because the current numbers are much worse (anywhere between 95% to 160%, with one case of 180%) than what they've propagated (180% across the board), and so one can think there is some flaw somewhere here, as well. (And there is indeed the case of L1D trashing, that they claim to be responsible for only 3% decrease.)

When you disable a cluster in BIOS, you do the same thing as AMD's diagram.

What diagram? Do you mean this? Which part of it?

What were doing
Core 0 - one cluster
Core 1 - disabled
Core 2 - one cluster
Core 3 - disabled

Do you mean, if we disable every other "core" in the BIOS? Then no, you will get this:
Core (Module) 0 - one cluster
Core (Module) 1 - one cluster
Core (Module) 2 - one cluster
Core (Module) 3 - one cluster

ps. perhaps the title of the thread should be changed to "Thread separation vs. turbo", or something like that, to be more meaningful.

[The Stilt]

Indeed there should be a update coming for Windows which optimizes Turbo Core functionality on Zambezi.
Currently Windows (7 atleast) is throwing the load from core to core which sometimes neutralizes the effect of the Turbo Core feature.
This is because the load is not being run on the currently boosted core(s).

But guys... please...

When talking about Zambezi please use the correct terms to avoid any further confusion.

A Zambezi node consists of: Four compute units and eight cores.
Each compute unit contains two cores.

In some of the slides a compute unit was called as a module, however thats not the official term.