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

[Oese]

hmm here 1/3/5/7 is faster at lower clocks...

now the question is, which setting is consuming less energy? In (one of) these tests?

Still, it was shown that in the end ressource sharing provides better perf/watt, given same clocks. Now, with less clocks on the non-sharing variant, idk if this would change? Or, due to higher voltage, its even more the case?

My thinking is still

1.) BD 8-core doesnt consume that much in situations where not all cores are loaded, or cores are hindering each other (whilst this would be a case where some hindrance might consume power itself, like cache misses and chache traffic due to this), therefore one should use bulldozer as it is and dont worry too much about consumption, its only bad in full-load-conditions and doesnt affect perf/watt in medium loaded situations

2.) BD running as 4-core 1/3/5/7 is more power efficient, due to less clocks for same Performance and disabled (not-needed-for-gaming) additional cores, whilst it is better then some 1/2/3/4 BD

3.) For Gaming workloads, anything then 4100 BD has too many cores and too much overhead, so go for this.

Background is, i'd like to buy a 8-core, cause this has the most opportunities and i dont like disabled chips (case is different if those can be enabled however ), and somewhere i think for megatasking 8 cores would be nice, if i could have good perf/watt in lower threaded/loaded situations.

The Win8 scheduler seems to provide this, either in 1/2/3/4 or in 1/3/5/7 configurations.

Again the question, chew*: Those 8-cores clock higher when you go for 1/2/3/4 compared to all cores enabled, or do they clock lower then all cores enabled with 1/3/5/7 ??

[dess]

if we overclock to the max with 2 CUs and 4 threads, vs 4 CUs and 4 threads, you can OC higher because heat wont be your issue nearly as quickly. we saw that in low threaded apps 5% gains are expected from not sharing CUs (in gaming for example), while he can get more than 5% higher clocks which will offset that value.

Well, it seems it's true only for 3DM2001...

[Manicdan]

Well, it seems it's true only for 3DM2001...

theres proof of that from a review showing MOST games they tested, not just chews test of synthetic gaming benchmarks.

[dess]

theres proof of that from a review showing MOST games they tested, not just chews test of synthetic gaming benchmarks.

Could you please provide a link for that?

Some games at least gain 10% this way. I'm not even sure if it's including (higher) turbo for the shared mode, so it could be more C2C.



amd_fx_win8_performance_boost.jpg

[chew*]

Sure, what about GT1 in vantage ( the only GT that matters is GT1 for cpu since gt2 is all about GPU and the diff for that can be bench run variance ), or another GT in 3d 11........

I can always add 05 and 03 gametests to the mix, they are indicative of 01......not surprising since they only benefit from 2 cores.

Bottom line this is not the holy grail fix due to the fact that there is a clock hit.

I haven't gotten to actual games yet but i can already see a trend, if not full loading 4 cores the advantage goes to clock speed..........

This test is far more real world due to fact that in real world clocks will vary betwen the overclocks.......running at same speed is pointless for a compare since headroom quite obviously varies enough to matter.

If the findings bother people it's not a big deal, this is stuff i'm looking at for my own daily rig, I can keep the info to myself.

I'll give 8 cores a shot later but i was pegging over 60C....even at my rather conservative settings.

[Manicdan]

Could you please provide a link for that?

Some games at least gain >10% this way. I'm not even sure if it's including (higher) turbo for the shared mode, so it could be more C2C.



amd_fx_win8_performance_boost.jpg

i think your mixing up statements and there were quite a few

A: using 1 thread per CU helps gaming performance, TRUE
B: using 2 threads per CU and overclocking farther can in many cases give more perf if your not on a super strong WC system, also TRUE

heres a link with details for 'A'
http://www.hardware.fr/articles/842-...acite-cmt.html
4/6 games got ~5% faster when using 1 thread per CU (but at the same clocks), the other 2 games were identical, aka no penalty.

the windows 8 bonus i hope is not just about making the turbo work better by ensuring threads stick to modules so the others can turn off and reduce TDP, since thats something we should have working now with very simple fixes rather than needed to purchase a brand new OS.

[mR Yellow]

Thanx for all the interesting info/tests. AMD chips, as usual, are always fun to play with.

I currently have a 1055T in my Sabertooth 990. You guys recon it's worth getting DB?

[EniGmA1987]

I really hope they come out with a patch for Windows 7 that includes some of the scheduler stuff Windows 8 has once it is all perfected. I like my Windows 7, I dont want to have to change to Win8 just for better performance on this CPU.

[dess]

i think your mixing up statements and there were quite a few

I don't think I'm mixing up anyting.

A: using 1 thread per CU helps gaming performance, TRUE
B: using 2 threads per CU and overclocking farther can in many cases give more perf if your not on a super strong WC system, also TRUE

C: Depending on the workload, sometimes it's A, sometimes it's B that gives you the higher performance...

heres a link with details for 'A'
http://www.hardware.fr/articles/842-...acite-cmt.html

Thanks, but I've inserted charts here from this article already.

4/6 games got ~5% faster when using 1 thread per CU (but at the same clocks), the other 2 games were identical, aka no penalty.

Nice, but it's just a few games.
Edit: I think you're reading the chart wrongly. All games gained some (2-12% [and 40% in case of Houdini 2, although it's not of typical game load]) with 4CU/2C, compared to 2CU/4C.

the windows 8 bonus i hope is not just about making the turbo work better by ensuring threads stick to modules so the others can turn off and reduce TDP

This diagram (and the explanation here) suggests that the default mode for unrelated threads is to execute them without sharing, if possible.

since thats something we should have working now with very simple fixes rather than needed to purchase a brand new OS.

What fixes do you mean?

Regarding utilizing cores with priority on unsharing (edit: I mean automatically), and the need for Win8: I've already pointed out it should be easy to do it on Win7 by simply enabling SMT-aware scheduling (if it's true it's already supports it in case of Intel CPU's with HT). Although, in some cases you will prefer to disable it.

[Manicdan]

I don't think I'm mixing up anyting.
C: Depending on the workload, sometimes it's A, sometimes it's B that gives you the higher performance...
Thanks, but I've inserted charts here from this article already.
Nice, but it's just a few games.

you just proved your confusing people by mixing statements

A is about not sharing resources, B is about using extra TDP to overclock/turbo higher, it can only be both if you leave the chip stock and try to play with schedulers to see the results. if you have 2 chips with the same clocks, then A will be better. but if you have 2 chips with the same TDP then B will be better. you cant say that C is either because they are opposite sides of the same coin.

and just a few games is the only games weve seen tested. so for now its 66% of games out there do notice an increase by not sharing resources IF CLOCKS ARE LEFT ALONE. the test with WoW by Toms Hardware mentioned no details about how that 10% was gained, and it looks to be based around 'B' with using higher turbo settings. but they dont tell us anything.

[son14]

i think your mixing up statements and there were quite a few

A: using 1 thread per CU helps gaming performance, TRUE
B: using 2 threads per CU and overclocking farther can in many cases give more perf if your not on a super strong WC system, also TRUE

heres a link with details for 'A'
http://www.hardware.fr/articles/842-...acite-cmt.html
4/6 games got ~5% faster when using 1 thread per CU (but at the same clocks), the other 2 games were identical, aka no penalty.

the windows 8 bonus i hope is not just about making the turbo work better by ensuring threads stick to modules so the others can turn off and reduce TDP, since thats something we should have working now with very simple fixes rather than needed to purchase a brand new OS.

This is so fascinating, that i had to register at this forum. I spent a lot of time this weekend, analyzing the french and the results from the beginning of this thread. And it seems that the performance gain of deactivating cores in the BIOS, is higher than that what the guys from France did. They see 2M/4C in sum 15 % behind 4M/4C. In the first post here it was reached 22 %. Just have a look at Fritz Chess. Here 8813/6335=139 % and they tested 8130/6417=127%. So if you try to reach 4M/4C(not gaming) from a base of 4 GHz, you have to clock the 2M/4C at 5 GHz. The power consumption would be significantly higher. This may change, when a native designed 2 computing unit chip for FX-4X00 is released.

The best effort in gaming may be deactivating 1 CU in bios, so you get a FX-6X00. And then take a piece of software, like "core affinity resident", for automaticaly binding the games to the right 3M/3T cores, simulating the windows 8 scheduler. Overclocking should be better than with 4M/4T and you get the advantage of 1 thread/CU.

I think that overclocking is not the best idea to gain more efficient performance, when you get a power draw of >500 watts. If you take a look at the effects of northbridge oc in games:
http://www.rage3d.com/reviews/cpu/am.../index.php?p=9
than i favour 4CU/4C or 3CU/3C with nb oc. That should only fill the saved watts between power draw of FX-8150 in standard configuration and 4CU/4C or 3CU/3C:
http://www.hardware.fr/medias/photos...IMG0033837.gif
so you get from 84 to 110 watts without turbo. You can see that the turbo in 2CU/4C mode is horrible inefficient. It raises the power draw from 60 -> 95 watt = 60 % and gains performance of 11 %. It is horrible.

So pleeeeeaaase, benchmark the 4CU/4C with overclocked nb and 3,6 or 4GHz!

The guys here:
http://www.hardocp.com/article/2011/...mance_review/5
got 7-10 % better results with DDR-1333 instead of all the reviewers with 1866. Memory bandwith is bad, but only WinRar is not amused about that. Drystone, Lame, WPrime, ... all with better results through all of the benchmarks.

If you imagine 5-10 % 4CU/4C + 5-10% nb oc + 5-10 % memory miracle = 15-30 %. Its worth a look to bring them together in benchmarking.

[dess]

you just proved your confusing people by mixing statements

No, I don't. I was saying in some cases it's unsharing, in other cases it's turbo that gives higher gains. It's not contradicion.

A is about not sharing resources, B is about using extra TDP to overclock/turbo higher, it can only be both if you leave the chip stock and try to play with schedulers to see the results. if you have 2 chips with the same clocks, then A will be better. but if you have 2 chips with the same TDP then B will be better. you cant say that C is either because they are opposite sides of the same coin.

I wasn't speaking about doing both (using 4CU/4C mode and max. turbo) at the same time. Although, disabling every other hw thread lowers power consumption, thus making some room for lifting the clock, as well. Also, you can have max. turbo with 2 CU's and executing a dual-threaded program with 1 thread/CU.

but if you have 2 chips with the same TDP then B will be better.

Prove it.

and just a few games is the only games weve seen tested. so for now its 66% of games out there do notice an increase by not sharing resources IF CLOCKS ARE LEFT ALONE.

Seems you've missed my edit: I think you're reading the chart wrongly. All games gained some (2-12% [and 40% in case of Houdini 2, although it's not of typical game load]) with 4CU/2C, compared to 2CU/4C.

the test with WoW by Toms Hardware mentioned no details about how that 10% was gained, and it looks to be based around 'B' with using higher turbo settings. but they dont tell us anything.

There are three indications (the above results with the games, the diagram and the explanation for it) that it's rather A. Or perhaps both (with 2 CU's).
Wouldn't they scale more similarly with turbo alone?

[Manicdan]

No, I don't. I was saying in some cases it's unsharing, in other cases it's turbo that gives higher gains. It's not contradicion.


I wasn't speaking about doing both (using 4CU/4C mode and max. turbo) at the same time. Although, disabling every other hw thread lowers power consumption, thus making some room for lifting the clock, as well. Also, you can have max. turbo with 2 CU's and executing a dual-threaded program with 1 thread/CU.


Prove it.



Seems you've missed my edit: I think you're reading the chart wrongly. All games gained some (2-12% [and 40% in case of Houdini 2, although it's not of typical game load]) with 4CU/2C, compared to 2CU/4C.


There are three indications (the above results with the games, the diagram and the explanation for it) that it's rather A. Or perhaps both (with 2 CU's).
Wouldn't they scale more similarly with turbo alone?

chew is proving 'B' and im taking his answer as being accurate.
for the games i was looking at 4/4 being faster than 4/8, i completely ignored 2/4 because thats comparing a 4100 to an imaginary 4/4 chip. the perspective i was trying to show with point 'A' was the INCREASE in performance by chopping up an 8100's cores. (i have no idea what houdini 2 is, and based on details it looked like a multi-threaded program rather than a game)

i still belive the WoW example by TH was done by turbo alone. that game uses less than 3 full cores, but has 3 distinct threads, and the 2 weaker ones in total are weaker than the first one alone. so its possible they got it down to one CU and just let it max out at 4.2ghz. but if left untouched it might have used up 3 CUs and stuck between 3.6 and 3.9ghz. that game is actually rather important to me since thuban isnt that great at it, and it is very cpu limited, but also very limited to single threaded perf. i really wish TH did more talking about the results and not just the theories. showing us which threads were parked for the game, or what the clock speed was in their test. all we know is that Win8 helps, but it can be quite a few combinations to make it work out best. i wonder if Win8 will be smart enough to test whether using a single CU for the game and max out turbo will be better than using 2 CUs where it MIGHT still be able to near max out, but also give the heaviest thread its own resources.

[dess]

chew is proving 'B' and im taking his answer as being accurate.

I think you should study his results more closer...
Which answer?

for the games i was looking at 4/4 being faster than 4/8, i completely ignored 2/4 because thats comparing a 4100 to an imaginary 4/4 chip.

What's the point in doing so in case of games with 4 threads or less?

Max. turbo is enabled only if there are 2 active CU's. So, if you want to know if it's either max. turbo or unsharing that gives you more performance with 2-4 threaded games, it's indeed 4CU/4C vs. 2CU/4C that you should take for a comparison...

the perspective i was trying to show with point 'A' was the INCREASE in performance by chopping up an 8100's cores.

The whole topic is all about this (for programs with 2-4 threads), why do you think you need to try to show this?

(i have no idea what houdini 2 is, and based on details it looked like a multi-threaded program rather than a game)

AFAIK it's a well-threaded chess game.

i still belive the WoW example by TH was done by turbo alone. that game uses less than 3 full cores, but has 3 distinct threads, and the 2 weaker ones in total are weaker than the first one alone. so its possible they got it down to one CU and just let it max out at 4.2ghz.

You can get max. turbo with 2 CU's.

There are two probable scenarios:
- Execute the more demanding thread on the first CU alone, the two weaker ones on the second CU, get max. turbo activated, and let the other CU's go to sleep.
- Execute all three threads alone in 3 CU's, at all-cores turbo.

Now, I don't think it runs even on Win7 at stock 3.6 GHz all the time... Given it's only 3 threads, it can run partially even at max. turbo there, only there is some penalty because of changing cores and so switching frequencies, etc. At least all-cores turbo can be active most of the time. Thus, 10% gain Win7 vs. Win8 can't come alone from turbo.

[chew*]

OK dirt 2 all same settings GPU wise.

4 threads on all systems.

Screenshot to show cpu usage of bench.



Sandy



4600 1/3/5/7



4800 1/2/3/4



Now to find a game that uses 2 threads.

[chew*]

Tested the 2 core theory out or "lightly threaded cpu app" now i just need a game to prove my point

4600 1/3/5/7


4800 1/2/3/4


The real kicker is heaven doesn't gain to much from CPU so this is definitely a significant gain from clock speed.

[son14]

Tested the 2 core theory out or "lightly threaded cpu app" now i just need a game to prove my point

4600 1/3/5/7


4800 1/2/3/4


The real kicker is heaven doesn't gain to much from CPU so this is definitely a significant gain from clock speed.

Try Crysis Warhead or Crysis. I prove CrWh on Intel i5 760@3,3 GHz and HD6950, 640x480 lowest and got:

2 cores 238 average
3 cores 243 average
4 cores 243 average

3,4 or less doesnt matter, but 2 cores got well busy during benchmark run.

[chew*]

HL2 lost coast is enough to prove my point.

4600 1/3/5/7


4800 1/2/3/4


And to show this game is impacted by clocks



Both the sandy rig and the 4800 BD run are score limited due to a 300fps cap.....

[EniGmA1987]

4600 1/3/5/7



4800 1/2/3/4

So it looks like in Dirt 2 that 4 modules with 1 thread each at a slower speed does much better than 2 modules using both cores at a higher speed

[chew*]

So it looks like in Dirt 2 that 4 modules with 1 thread each at a slower speed does much better than 2 modules using both cores at a higher speed

Yep basically I came to the conclusion that any game that uses 4 cores heavily ( like 75% or better ) will benefit from 4core 4 cluster configs........problem is the clock hit it takes.

Preety much any game that is a console port (75% games made for pc? ) that will only 50% load 4 cores or load 2 cores heavily will benefit from a 2core 4 cluster config due to the higher clocks attainable in that configuration.

I've tested like 23 chips like this, all of them take a clock hit in the 4core 4 cluster config, some much worse than others.

[Manicdan]

snip

the idea of comparing 4/8 with 4/4 instead of with 2/4 is to isolate things to a single detail and test the crap out of that detail to understand it. if you add in clock rates or TDP then your looking at a scenario closer to reality, but not with scientific precision.

what chew is doing is great, but he would need to do about a dozen more tests with each scenario to cover everything. so instead he built a few real life examples and showed when either is better. i would just like to see if he can set up a turbo on the 4/4 to let it get a little higher clock rate when its a low threaded game. he might only get an extra 100mhz, but it might be enough to make it better in many more cases. i havnt seen enough examples of how turbo can be set up with overclocks and how well its working once set up.

[chew*]

the idea of comparing 4/8 with 4/4 instead of with 2/4 is to isolate things to a single detail and test the crap out of that detail to understand it. if you add in clock rates or TDP then your looking at a scenario closer to reality, but not with scientific precision.

what chew is doing is great, but he would need to do about a dozen more tests with each scenario to cover everything. so instead he built a few real life examples and showed when either is better. i would just like to see if he can set up a turbo on the 4/4 to let it get a little higher clock rate when its a low threaded game. he might only get an extra 100mhz, but it might be enough to make it better in many more cases. i havnt seen enough examples of how turbo can be set up with overclocks and how well its working once set up.

I think it would just be easier to set up 2 profiles in bios, load whatever one will benefit you most for the app your running. Thats all i'm doing now.

As far as 8 core operation I tried 4.6 on this air cooler at this voltage, no dice.

[dess]

Yep basically I came to the conclusion that any game that uses 4 cores heavily ( like 75% or better ) will benefit from 4core 4 cluster configs........problem is the clock hit it takes.

Why is it a problem for them if those are still faster this way than with the higher OC?

the idea of comparing 4/8 with 4/4 instead of with 2/4 is to isolate things to a single detail and test the crap out of that detail to understand it. if you add in clock rates or TDP then your looking at a scenario closer to reality, but not with scientific precision.

Yes, this way you understand that an up to 4-threaded app won't scale any further with more cores than 4. What a revelation...
And you will learn nothing about CMT scaling (in case of up to 4-threaded apps, again).

what chew is doing is great, but he would need to do about a dozen more tests with each scenario to cover everything. so instead he built a few real life examples and showed when either is better. i would just like to see if he can set up a turbo on the 4/4 to let it get a little higher clock rate when its a low threaded game. he might only get an extra 100mhz, but it might be enough to make it better in many more cases.

Do you mean, to see if an even higher clock is possible in 4CU/4C mode, but with 2CU's gone to sleep? Could be.

i havnt seen enough examples of how turbo can be set up with overclocks and how well its working once set up.

I think it can be done by modifying the VID and FID values that are assigned to the various power states.

I think it would just be easier to set up 2 profiles in bios, load whatever one will benefit you most for the app your running. Thats all i'm doing now.

How great it would be if it all worked full automatic... Could it made work under Windows with some driver? Or only MS could do it?
Well, in times like this Linux would come handy...

As far as 8 core operation I tried 4.6 on this air cooler at this voltage, no dice.

What if you enable only two cores in two CU's...?

[verndewd]

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

This almost supports the ex engineers whine about AMD not personally crafting the initial chips.

[Oese]

maybe clock hit depends on sth like bootstrap cu need 2 cores?

Like enable first cu fully, rest only one cu/c

sth like this