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

What I'm about to deal with here is comparing 2CU/4C and 4CU/4C Bulldozers.
(CU stands for Compute Unit, or equivalently 'Module')
It can be an excuse for Bulldozer's initial poor single-thread performance:
benchmark tools are just under-optimized for that kind of architecture, but not virtually a 'poor' performance if some optimization is done.

Here's the test scheme: I'm sure that everybody understand what these pics meaning for.

Attachment 121155
(4M/8C)

Attachment 121156
(real die: do they seems likely?)

Attachment 121157
(2M/4C: traditional manner for AMD)

Attachment 121158
(4M/4C: conceptual diagram)

But there's a flaw: current version of CH5F BIOS doesn's support core-by-core on/off. Am I finished then?

Attachment 121159

...Not really. Newest version of BIOS supports such function!

Attachment 121160
Attachment 121161

Here's the result:

Attachment 121162
Attachment 121163
Attachment 121164
Attachment 121165
Attachment 121166

▲ (there are six more results. I'll attach a reply to this thread)

(continued from above)

Attachment 121167
Attachment 121168
Attachment 121169
Attachment 121170
Attachment 121171
Attachment 121172

▲ See what I mean? ;-)

so for anyone wonder either, what would a 4M/4C chip be like, and how much scaling is lost sharing resources, you've answered both,
seriously thank you for having what is so far, the best review.

I appreciate your efforts. However, it's likely that you're largely seeing the performance hit of the cache thrashing issue recently discovered since disabling a core in each module would prevent that contention.

Excellent review, DGLee. Could you try 4CU/4C vs 4CU/8C with 2 and 4 threaded applications? I would like to see if having the cores active impact performance in those scenarios, to see if: a) Windows scheduler is at fault and can be optimized (you could force the threads to run on a specific CPU) and b) if having the cores "online" but inactive impacts performance.

I questioned this.. a long time ago.

The answer from AMD was it was better off in the '2cu 4c' arrangment, and turboing higher.. however your results seem to disagree with that quite a bit.

any chance you can test single thread without affinity set between the two configs? I'd like to see the implications of threads bouncing between modules vs cores

Quote:

---

Originally Posted by mAJORD

I questioned this.. a long time ago.

The answer from AMD was it was better off in the '2cu 4c' arrangment, and turboing higher.. however your results seem to disagree with that quite a bit.

any chance you can test single thread without affinity set between the two configs? I'd like to see the implications of threads bouncing between modules vs cores

---

very interesting results.

can you also pls try to kick that NB to higher regions and see how it reacts?

I understand you are talking about disabling alternating cores so each core unit only has one core activated. but your graphs confuse me. you will have to explain what your "4cu/8c" notation means, what each tested configuration means, and what your conclusion is based on this chart.

basically he is talking about using a module as a single core, single thread....
4 modules with a total of 4 threads is obviously faster than 2 modules with a total of 4 threads, since the 4 threads from 2 modules will be sharing resources while the other option don't...

when using less than 4 threads the optimal choice would be to use it like this (1t per module), but probably windows cannot make the distinction?
and that's what is going to be fixed for Windows 8?

just for :banana::banana::banana::banana:s a giggles can you run a single thread benchmark like Pi.

Quote:

---

Originally Posted by bamtan2

I understand you are talking about disabling alternating cores so each core unit only has one core activated. but your graphs confuse me. you will have to explain what your "4cu/8c" notation means, what each tested configuration means, and what your conclusion is based on this chart.

---

CU = Computational Unit

AMD's internal tech term for PR "Module"

I'm a bit puzzled by what's the news here. You're basically proven an axiom : in any CPU were you have resource sharing among 2 threads, running only one ( thus giving it the whole resources ) it will run better.
If you take a Core 2 and disable 1 core, rest assured, the performance of that one thread will be better than running the same thread with both cores active.

The whole point of AMD's aproach is to avoid exactly that : don't make a fat core ( what you're suggesting ), but skinny ones and lots of them. On desktop, as BD has proven, this is a failure.

AMD could have created BD as a 4 core with each module being transformed in a fat core, but that's SB reloaded.

Quote:

---

Originally Posted by savantu

I'm a bit puzzled by what's the news here. You're basically proven an axiom : in any CPU were you have resource sharing among 2 threads, running only one ( thus giving it the whole resources ) it will run better.
If you take a Core 2 and disable 1 core, rest assured, the performance of that one thread will be better than running the same thread with both cores active.

The whole point of AMD's aproach is to avoid exactly that : don't make a fat core ( what you're suggesting ), but skinny ones and lots of them. On desktop, as BD has proven, this is a failure.

AMD could have created BD as a 4 core with each module being transformed in a fat core, but that's SB reloaded.

---

Thanks, but I think we all know that :p . This is not the news section.. So no ones claiming it's news.

Some people are interested in the effects of the shared resources, DGlee has put the effort in to show the results, that's all there is to it.

i wonder that the next bios of asus ..will it have more improve

What you actually proved is the opposite (for now), that AMD gets 33-59% on the CMT :)

Chess 11800/8813=1.3389 ?
Wprime 13.814/9.531=1.4494
Winrar 4467/3027=1.4757
3d06 5803/4134=1.4037
3dvantage 19215/12102=1.5878
3d11 6340/4289=1.4782
CB R10 20552/15033=1.3671
CB R11.5 6/3.8=1.5789
Blender 9.76/7.16=1.3631
X264 37.23/25.18=1.4786
Transcode (222+210)/(185+135)=1.35

well, unless you run single thread program like SuperPi however (for comparing Single 'Core' Module and Dual 'Core' Module), then you can see if there is an impact to single threaded stuff :)

time to finetune the win7 os and create an application that shoots newly started applications to different modules/cores eliminating the usage of the second module as much as possible untill one need 5 cores :)

btw patch for win7 shoudln't be taht hard, Vmware already got it in there VSphere esxi scheduler.

Quote:

---

Originally Posted by Particle

I appreciate your efforts. However, it's likely that you're largely seeing the performance hit of the cache thrashing issue recently discovered since disabling a core in each module would prevent that contention.

---

question would not the trace cache prevent that ?
it's 4Kbytes pre core. 8 way association.

Quote:

---

Originally Posted by ThePointer

What you actually proved is the opposite (for now), that AMD gets 33-59% on the CMT :)

Chess 11800/8813=1.3389 ?
Wprime 13.814/9.531=1.4494
Winrar 4467/3027=1.4757
3d06 5803/4134=1.4037
3dvantage 19215/12102=1.5878
3d11 6340/4289=1.4782
CB R10 20552/15033=1.3671
CB R11.5 6/3.8=1.5789
Blender 9.76/7.16=1.3631
X264 37.23/25.18=1.4786
Transcode (222+210)/(185+135)=1.35

well, unless you run single thread program like SuperPi however (for comparing Single 'Core' Module and Dual 'Core' Module), then you can see if there is an impact to single threaded stuff :)

---

I think IPC would be an improvement over Phenom II if that is the case...
35% of 20s would be 7 seconds, that would put stock 4.2 Ghz around 13 seconds :O
Currently taking 6.3 Ghz to do 13 seconds from what I see.

Although nice to see somone tested it this is really nothing more than what i stated in the BD info thread.

Killing resource sharing improves performance.

There are other things that can be done as well.

The next logical step is this.

Run max 24/7 clocks for deneb, thuban, and BD then compare performance.

That should answer questions for faithfull AMD users if BD is an upgrade for them.

Quote:

---

Originally Posted by chew*

Although nice to see somone tested it this is really nothing more than what i stated in the BD info thread.

Killing resource sharing improves performance.

There are other things that can be done as well.

The next logical step is this.

Run max 24/7 clocks for deneb, thuban, and BD then compare performance.

That should answer questions for faithfull AMD users if BD is an upgrade for them.

---

So if I wanted less heat output, lower power consumption, higher theoretical clocks and much better single thread performance I'd kill half the cores in bios and overclock the CPU to hell and back, right?

In my opinion the CPU wouldn't look half as bad if ST performance gained 20% from the non-resource sharing up to 4 threads...

Quote:

---

Originally Posted by BeepBeep2

So if I wanted less heat output, lower power consumption, higher theoretical clocks and much better single thread performance I'd kill half the cores in bios and overclock the CPU to hell and back, right?

---

Cores?

You can not kill cores, get that marketing :banana::banana::banana::banana: out of your head.

You can disable clusters inside cores.....

There is no mention of the term module in patents so lets bury the damn term......................

Me personally I am going to run 1 in a daily rig with all 4 cores but clusters B disabled in all cores.

Run it at 4.8-5.0 24/7 with high ram and call it a day.

Quote:

---

Originally Posted by chew*

Cores?

You can not kill cores, get that marketing :banana::banana::banana::banana: out of your head.

You can disable clusters inside cores.....

There is no mention of the term module in patents so lets bury the damn term......................

---

Sorry, the bios says cores. ;)
Sorry, "clusters inside the computational units"

Why the hell couldn't AMD figure out a way to disable this resource sharing when 4 or less threads are executed?

Quote:

---

Originally Posted by BeepBeep2

Why the hell couldn't AMD figure out a way to disable this resource sharing when 4 or less threads are executed?

---

good question, one I don't have an answer for at this point in time.

So if i disabled 4 clusters, will it run games better?

Quote:

---

Originally Posted by chew*

Cores?

You can not kill cores, get that marketing :banana::banana::banana::banana: out of your head.

You can disable clusters inside cores.....

There is no mention of the term module in patents so lets bury the damn term......................

Me personally I am going to run 1 in a daily rig with all 4 cores but clusters B disabled in all cores.

Run it at 4.8-5.0 24/7 with high ram and call it a day.

---

@ Chew ,,Exactly +1,,

It will be a fun one to play with regardless...its just the beginning my friends.