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

[The Stilt]

Like enable first cu fully, rest only one cu/c
sth like this

Thats mixing and it is not possible

[EniGmA1987]

Is it not possible just because the bios isnt there yet or because it actually isnt supported to disable one core on the rest from the processor itself?

[The Stilt]

Is it not possible just because the bios isnt there yet or because it actually isnt supported to disable one core on the rest from the processor itself?

It is a design limitation as far as I know.
You can run the compute units either in single or dual core mode but not at the same time.

[verndewd]

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

maybe they need to seperate l3 into 4 segments of two cores each. maybe the atx form factor is reaching its limits and intel need to redesign the damn thing for 8+ cores

[verndewd]

It is a design limitation as far as I know.
You can run the compute units either in single or dual core mode but not at the same time.

but is it a platform limitation? lets face it its the same freakin platform as the 90's bs intel atx. maybe the entire platform needs to evolve.

[omninmo]

great work chew, this may not be what everyone was hoping to hear but it will surely be of assistance to people reading up to see whther BD fits their needs or not!

So on to the juicy stuff, your findings are consistent with a classic resource sharing bottleneck, or so it seems..
To sum it up:
- when a CPU heavy task fully loads both clusters in a CU, the shared resources have a hard time feeding or dealing with both clusters so you notice a bottleneck and performance per thread drops accordingly -> ergo, 4c/4cu wins
- when only one cluster is loaded, or both clusters are being used, but one of them (or both) is/are not being fully loaded, shared resources can keep up and so performance per thread takes a lesser/no impact -> ergo, clockspeed wins

This is fully understandable and seems perfectly logical

Next subject, max clocks:
I'm still confused by something here, please see which interpretation is the correct one (or if both apply):
- 4c/2cu has lower power/thermal footprint, ergo for the same voltage, same cooling, it will clock higher?
or
- Luck of the draw means that you're likely to get different OC potential on all modules, and therefore you might (probably) end up with having to choose a cluster from a module with worse potential when taking the 4c/4cu approach, therefore limiting overall clocks a bit?

Also, for this last part, do you think it plausible that with better air or water cooling (for instance, i have 360 rad custom water loop) and maybe some extra volts one could push those limits higher and reach higher clocks? Perhaps 4.8 ~ 5Ghz across all cores? or is this a pipe dream for this stepping?

Last question:
Have you any idea of the actual power consumption difference between running 4c/2cu vs 4c/4cu vs 8c/4cu @ same clocks all around, for instance, 4.6Ghz? I am considering an upgrade but at this point im afraid my TX650 would not cut it powering a BD 8c OC plus a 6970 or 6870 CF?

My personal opinion on this:
Best case scenario (and the most balanced option for varied daily usage) would be for us to be able to leave them all on, and have AOD or something clock each loaded core higher.. While at the same time having some way to make the os treat one of the clusters in each cu as a "Hyperthreading-like" affair... However im guessing this would require a hefty power draw and adequate high end cooling to control..

Crappy math:
So taking into account an announced (and somewhat accurate, from what we can see) performance penalty of roughly 20% when fully loading both clusters (ex: Civilization or starcraft) we can speculate that it would take roughly 5 ~ 5.2Ghz clocks all around for a "FX-4110" to match performance of a PII x4 @ 4 ~4.2 Ghz. Whereas for less intense games with a 2main, 2 secondary threads would require only around 4.3 to 4.4Ghz.

This would be an avantage to the octa-cores that they could "shine" in both scenarios, we would just need a way to make this hassle free and be able to pair this up with NOT having to reboot or manually change these settings whenever we want to game, or encode, or whatever :P

once again, thank you guys for your awesome work! and also everyone sorry for the long post, feel free to ignore if you wish but i just wanted to see if I had all the facts straight from this "investigation"

Please, do correct my conclusions if you seem fit as I have a feeling these are some of the points other people could benefit from aswell

[son14]

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.

The answer could be simple. If you use 2CU/4C, there are 2 floating point units available, in 4CU/4C mode 4. So a task using 4 cores, gets access to 4 vs. 2 fpus. The performance gain in 50% int and 50 % fpu scenarios, is up to 25 %. Xbox ported games, mostly use 2 cores. So they get access to 2 fpus in both scenarios and you see performance gain of about 5 % only, from doubled cache per core.

Console ports are not the problem, there is always enough power for a FX6 or FX8 to reach >60 fps. The problem is Dirt, F1, Civ5, Crysis 2,... . Northbridge overclocking and 4CU/4C can be the difference between playable or not.

[Oese]

i just installed win 8 beta, on a thuban system though. It seems, its stable enough to work for some gaming, and i guess this scheduler is more aware of what BD needs. I am looking forward to tests coming in at the other thread covering this..

[wez]

Last question:
Have you any idea of the actual power consumption difference between running 4c/2cu vs 4c/4cu vs 8c/4cu @ same clocks all around, for instance, 4.6Ghz? I am considering an upgrade but at this point im afraid my TX650 would not cut it powering a BD 8c OC plus a 6970 or 6870 CF?

I'm over 650W total load with a oc'ed 1090t and two 6950's @ 6970. So I'd say single GPU only, with that PSU.

[son14]

great work chew, this may not be what everyone was hoping to hear but it will surely be of assistance to people reading up to see whther BD fits their needs or not!


Have you any idea of the actual power consumption difference between running 4c/2cu vs 4c/4cu vs 8c/4cu @ same clocks all around, for instance, 4.6Ghz? I am considering an upgrade but at this point im afraid my TX650 would not cut it powering a BD 8c OC plus a 6970 or 6870 CF?

Take a look here at the picture "Fritz Chess benchmark 4.3 (watts atx 12v)":
http://www.hardware.fr/articles/842-...windows-8.html

2/4 = 60 watts (100%)
4/4 = 84 watts (140 %)
2/4 turbo = 95 watts (158 %)
8/8 = 101 watts (168 %)

totally power consumption in your scenario (oc and with HD6990) is about 500 watts in peak:
http://www.rage3d.com/reviews/cpu/am.../index.php?p=7

To reach 4/4 @stock in dirt with 2/4, you nearly have to double the power draw. In other games (console ports) its useless, because fps are mostly >60 fps and there is no gain in performance, gpu is limiting.

There are two different methods to get what you want:

1. let the FX-8 or FX-6 work in stock mode and bind the game/the app manually through the task manager to 1357 or 0246 or use a tool like "amd overdrive" or "core affinity resident" to build a profile; you do not need the expensive asus board and simulate the windows 8 task manager in a smarter way than inefficient cu parking+turbo
2. get the FX-8 or FX-6 into 4/4 or 3/3 mode to build X4 oder X3; asus is needed; thats it but you have to go into bios every time

[omninmo]

Oh, and by the way folks... but this might be a decent solution to allow us to manage our threads the way we like them and give us the extra flexibility of deciding if we want 4c/2cu, 4c/4cu or 8c/4cu on an app to app basis WITHOUT much hassle and without having to reboot!

http://img.tomshardware.com/us/2004/...taskassign.zip

just set your turbo OCs accordingly and setup each game in the way it will benefit you the most? win/win?

if someone could please test this and see if it plays nice with BD and let the rest of us now, would be grateful!

EDIT: hmmm seems people are already aware of utilities like this, my bad xD

[son14]

http://www.hardware.fr/articles/842-...windows-8.html
1. They see 2M/4C in sum 15 % behind 4M/4C and fritz chess 8130/6417=127%

2. In the first post here it was reached 122 % in all tests and 8813/6335=139 % in fritz chess; fritz chess is 108 % of 1.

So 1. should be task manager method and 2. deactivating cores in bios.

Interesting detail:

you need 6CU/6C clock by clock to beat an FX-8150 through all benchmarks

[mAJORD]

awesome, chew!.. Only thing missing is a current deneb OC'd under same cooling etc (that i'd like to see)

[dess]

The answer could be simple.

This answer is not for that question, because he was speaking about 4CU/8C vs. 4CU/4C, which is irrelevant in case of games with 4 threads or less.
So, let's we go back to 4CU/4C vs. 2CU/4C.

If you use 2CU/4C, there are 2 floating point units available, in 4CU/4C mode 4. So a task using 4 cores, gets access to 4 vs. 2 fpus. The performance gain in 50% int and 50 % fpu scenarios, is up to 25 %. Xbox ported games, mostly use 2 cores. So they get access to 2 fpus in both scenarios and you see performance gain of about 5 % only, from doubled cache per core.

You're searcing for the answer in the right direction, floating-point wise. Although, let's not forget this FPU is 2-way SMT, with double the resources of K10's FPU: it's 2xFMAC vs. 1xFADD+1xFMUL, FMAC being 1xFADD+1xFMUL combined. The thing is that all of this computing power can only be fully utilized with FMA code, because you can't have an FADD and FMUL independently started in the same cycle on an FMAC unit, only if it's an FMA instruction.

Thus, for legacy code 4CU/4C mode means double the resources utilizable all the time (not shared with the second thread), per hw thread, so it can have independent FADD's and FMUL's started in the same cycle on the two FMAC units. (Extra bonus is that it can have also 2xFADD or 2xFMUL, not just 1xFADD+1xFMUL, like on K10.) In 2CU/4C mode, it can have usually only 1xFADD OR 1xFMUL started in the same cycle (mostly one FMAC unit available because of the second thread engage the other one), per hw thread.

(And so, I think there won't be such a significant gain in performance for FMA heavy code, going 4CU/4C from 2CU/4C. EDIT: Or, even if there be so, both cases will be signicicantly faster than the same algorithm with legacy code.)

...So a task using 4 cores...

(I think it's not the best wording, because less than quad-threaded apps can also be "using 4 cores" upon the constant core-variation of Windows. So, it's about how many threads an application have.)

The problem is Dirt, F1, Civ5, Crysis 2,... . Northbridge overclocking and 4CU/4C can be the difference between playable or not.

Definitely worth a testing.

It is a design limitation as far as I know.
You can run the compute units either in single or dual core mode but not at the same time.

So, you're saying if we enable the 2nd core here, then we can't disable ony other cores, alone in a CU?

Oh, and by the way folks... but this might be a decent solution to allow us to manage our threads the way we like them and give us the extra flexibility of deciding if we want 4c/2cu, 4c/4cu or 8c/4cu on an app to app basis WITHOUT much hassle and without having to reboot!

http://img.tomshardware.com/us/2004/...taskassign.zip

just set your turbo OCs accordingly and setup each game in the way it will benefit you the most? win/win?

if someone could please test this and see if it plays nice with BD and let the rest of us now, would be grateful!

Wow, I've forgot about this little utility. Perhaps it could be enhanced and adapted to BD, so all of it could work fully automatic! Would be even more useful!

[The Stilt]

So, you're saying if we enable the 2nd core here, then we can't disable ony other cores, alone in a CU?

If core 2 is enabled the CU0 would be running in dual core mode.
This means all of the remaining CU must be running in dual too.
You can disable more cores while core 2 is enabled but then you would need to turn a whole compute unit off.

I think the limitation is because each core needs to have a pair.
Atleast the schematics show all of the individual cores connected. I might be wrong thou.

[The Stilt]

So, you're saying if we enable the 2nd core here, then we can't disable ony other cores, alone in a CU?

If core 2 is enabled the CU0 would be running in dual core mode.
This means all of the remaining CU must be running in dual too.
You can disable more cores while core 2 is enabled but then you would need to turn a whole compute unit off.

I think the limitation is because each core needs to have a pair.
Atleast the schematics show all of the individual cores connected. I might be wrong thou.

[demonkevy666]

Take a look here at the picture "Fritz Chess benchmark 4.3 (watts atx 12v)":
http://www.hardware.fr/articles/842-...windows-8.html

2/4 = 60 watts (100%)
4/4 = 84 watts (140 %)
2/4 turbo = 95 watts (158 %)
8/8 = 101 watts (168 %)

totally power consumption in your scenario (oc and with HD6990) is about 500 watts in peak:
http://www.rage3d.com/reviews/cpu/am.../index.php?p=7

To reach 4/4 @stock in dirt with 2/4, you nearly have to double the power draw. In other games (console ports) its useless, because fps are mostly >60 fps and there is no gain in performance, gpu is limiting.

There are two different methods to get what you want:

1. let the FX-8 or FX-6 work in stock mode and bind the game/the app manually through the task manager to 1357 or 0246 or use a tool like "amd overdrive" or "core affinity resident" to build a profile; you do not need the expensive asus board and simulate the windows 8 task manager in a smarter way than inefficient cu parking+turbo
2. get the FX-8 or FX-6 into 4/4 or 3/3 mode to build X4 oder X3; asus is needed; thats it but you have to go into bios every time

the GPU they used alone used 375 watts with out OC ? 545 - 375 = 125 watts

[The Stilt]

the GPU they used alone used 375 watts with out OC ? 545 - 375 = 125 watts

CPU PWM efficiency is around 80% + 80-90% for the PSU, so no.

[PatRaceTin]

interesting thread here

[spurious]

Please, please, please tell me all (or at least some) of this testing is with 64-bit applications on a 64-bit OS.

The application I am working on, the 64-bit version is 40% faster than the 32-bit version. There are twice as many general purpose registers in 64-bit mode. My application makes heavy use of 64-bit integers and no floating point. It is like Fritz except it is a breadth first search instead of a depth first search. It will use all the memory that you can throw at it and thrash it hard.

It wouldn't be that hard to add a dialog box to let you set the threads to run on the desired cores.

My next machine was going to be a dual socket C32 machine, I need the memory slots more than I need processing cores. But after reading Anands "Rendering and HPC Benchmark Session Using Our Best Servers", I have concerns about the memory performance of BD in dual socket boards.

[radaja]

@chew*,thanks for all your testing on this matter,you sir are awesome for going out of your way to help the community

[chew*]

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

It would probably do well as a 6 core, the third core pair on this chip appears to be the really sore thumb on this chip.

Thats what always drops out in prime 95.

I gamed till like 5am this morning at 4.6 4 core, tbh I can not tell the diff between this rig or my sandy rig.

Both run very smooth with no glitches.

I did have some hitches with a kingston SSD though, not agreeing with the ch5 to well, might be firmware.

It's fine with a normal sata drive.

[chew*]

Please, please, please tell me all (or at least some) of this testing is with 64-bit applications on a 64-bit OS.

All my tests were done with 16g of ram installed, that should answer that question for you

I'd like to apologize also, my cpu-z is outdated, this is a retail chip older versions call it ES, just installed new version

[mR Yellow]

Once again thanx for all the usual posts and info. So refreshing not having to hear about BD bashing.
Kudos to Chew and the other xtreme addicts!

[Swanie]

It would probably do well as a 6 core, the third core pair on this chip appears to be the really sore thumb on this chip.

Thats what always drops out in prime 95.

I gamed till like 5am this morning at 4.6 4 core, tbh I can not tell the diff between this rig or my sandy rig.

Both run very smooth with no glitches.

I did have some hitches with a kingston SSD though, not agreeing with the ch5 to well, might be firmware.

It's fine with a normal sata drive.

Thanks Chew for all your testing and feedback. It nice to see you don't feel any real world difference between BD and SB. Looking forward playing with my BD once it arrives. Got my UD7 today