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

[chew*]

We are talking well over a 300 mhz variance clock hit with a 4cu/4core versus a hypothetical 4170 when it comes to stability for 24/7 useage.

Preety much we are almost talking deneb/thuban type 24/7 speeds making a 4cu/4core a moot point, if thats the case just drop a 955 in a 790 board.

[Oese]

you are referring to a theoretical stock part that needs to meet TDP, right?

how is the overclocking affected? I still cant believe without specific confirmation that oc-ability goes down with deactivated clusters, maybe in comparison to a 4170 it remains lower due to the more frontend logic and thus heat involved, and thats why it doesnt pay off?

well finally, these tests indeed show that perf/watt is better when sharing ressources, performance doesnt go down as much as the wattage decreases, so 4170 > 4C/4CU.

Question for me is, in lightly threaded workloads and with some working scheduler, I think we dont have these high power requirements? I mean the old thing, max power usage is compared with medium-threaded, hindered performance, whilst this architecture doesnt scale anywhere lineally in that aspect, so this doesnt make sense...

[STEvil]

Could it be that the "deactivation" allows the threads to use the resources of the CU more efficiently (turning off a cluster may not turn off its resources, just the "thread" showing up as a core to windows) thus generating slightly more power usage?

[chew*]

Could it be that the "deactivation" allows the threads to use the resources of the CU more efficiently (turning off a cluster may not turn off its resources, just the "thread" showing up as a core to windows) thus generating slightly more power usage?

My take on it is this, BD is a highly unbalanced chip so sure you might get 2 core 4 clusters to do 4.8 24/7 but the chance of running 4 core 4 cluster at 4.8 is slim to none. To do that we would need full cluster control and be able to deactivate cluster a and use cluster B if we wanted which currently is not possible.

I tested down to 4.5 4core 4 cluster and it was still failing prime versus 2 core 4 cluster doing 4.8 no issues..........could also be a bug in bios somewhere but atm that is wishfull thinking.

[Oese]

Air cooling? so the "8core" variants dont overclock high+stable anyway? 6100 ihave seen 4.8ghz 5ghz turbo already..

[chew*]

Air cooling? so the "8core" variants dont overclock high+stable anyway? 6100 ihave seen 4.8ghz 5ghz turbo already..

Yes air cooling.

I may be testing a tad diff than most.

Key settings are turbo off, HPC mode enabled, APM disabled and in power section PWM thermalling disabled.

Those settings will keep the CPU at full throttle 100% with no performance loss.

That said seeing 4.8 with 5 turbo without knowing all settings may not be what you think it is.

Also one of the key elements is when i shoot off these figures we are talking prime 95.

I can bench pifast at 5.2+........

Also heat is not an issue in my testing currently, with a heatsink swap i'm barely hitting 50C loaded so enabling other units is not heat related failure.

[The Stilt]

My take on it is this, BD is a highly unbalanced chip so sure you might get 2 core 4 clusters to do 4.8 24/7 but the chance of running 4 core 4 cluster at 4.8 is slim to none. To do that we would need full cluster control and be able to deactivate cluster a and use cluster B if we wanted which currently is not possible.

I tested down to 4.5 4core 4 cluster and it was still failing prime versus 2 core 4 cluster doing 4.8 no issues..........could also be a bug in bios somewhere but atm that is wishfull thinking.

You mean ability to disable a specific compute unit?
Or "cherry picking" the individual cores which OC best?

The boot strap CU can be changed by bios programming so in case the CU0 is bad it can be technically disabled.

[Oese]

Chew, alright thank you! what is this hpc option about anyway?

[chew*]

You mean ability to disable a specific compute unit?
Or "cherry picking" the individual cores which OC best?

The boot strap CU can be changed by bios programming so in case the CU0 is bad it can be technically disabled.

You got the right idea

Currently although bios calls it core 1-8 even though reality is 0-7.

Anyway if you disable cu 3 cu 4 is auto disabled. What would be a real treat is to be able to disable cu 3 yet be able to use cu 4 instead and so on. I can't gurantee that that will alleviate what I am seeing but never know until you try.

Alright thank you! what is this hpc option about anyway?

Best way to explain it, vantage scores enabled can jump 500 points cpu score wise.

[The Stilt]

You got the right idea

Currently although bios calls it core 1-8 even though reality is 0-7.

Anyway if you disable cu 3 cu 4 is auto disabled. What would be a real treat is to be able to disable cu 3 yet be able to use cu 4 instead and so on. I can't gurantee that that will alleviate what I am seeing but never know until you try.

Try C5F 9911 bios.
I asked Asus to add control for individual cores.
It works perfectly and you can do the 'cherry picking' (i.e disable CU2 and leave the rest active).
The only limitations are: You cannot mix compute units working in single & dual core mode or change the boot strap core (so hope your CU0 is good)

[chew*]

Try C5F 9911 bios.
I asked Asus to add control for individual cores.
It works perfectly and you can do the 'cherry picking' (i.e disable CU2 and leave the rest active).
The only limitations are: You cannot mix compute units working in single & dual core mode or change the boot strap core (so hope your CU0 is good)

I have it, thats what i was testing with.

Problem is if core 2 = disabled then core 3 must be disabled, and so on.

I know core 1 is always enabled and can not be disabled, thats not so big a deal though as is being able to control the rest.

I had 0062 (where we had asus first implement this ) back when we did Lhe which is why we had this added but it gained us nothing and ended up setting record on 0057 instead.

It's why in the BD info thread i suggested that not everyone had acess to all pertinent info and or bios to gain the max potential out of BD

Basically what i would like to try is this. If we speak on what the bios terms are 1-8.

1 is auto enabled and can not be disabled, thats fine. currently i can only do this 1/3/5/7.

What i want to try is this. 1/4/6/8 or any combo pair if possible.

in my current setup 1-4 is fine but 5 and 7 are weak so want to try 6/8

[The Stilt]

Yeah, thats a design limitation as far as I know.
The "slave" cores (2, 4, 6, 8) cannot work without the master core enabled, however I am not sure about this.

So let the hunt begin

[chew*]

Yeah, thats a design limitation as far as I know.
The "slave" cores (2, 4, 6, 8) cannot work without the master core enabled, however I am not sure about this.

So let the hunt begin

As of now i concede the hunt to someone else.

If i have time i will test through chips one last time but 23 chips is exhausting.

As for hunting through 23 chips i have 5 candidates that should do 5 gig prime 95 once on water with cores 1-4. They all do 4.8 sub 1.4v currently, some 4.9.

I guess I will have to retest them all for core 1/3/5/7 but it was looking grim even at 4.7 that way.

So far I still say 1.30 is the sweet spot vid

I tested 4 recent chips and clocks increased greatly as i went up in vid from 1.25 to 1.30

The 1.25 chips could not even post at 4.8 sub 1.4v

[dess]

We are talking well over a 300 mhz variance clock hit with a 4cu/4core versus a hypothetical 4170 when it comes to stability for 24/7 useage.

300 MHz is 6% of 5 GHz, while the gain in performance on 4CU/4C is up to 26% C2C! How is it not worth it?

Problem is if core 2 = disabled then core 3 must be disabled, and so on.

Do you mean you cannot enable core 3, even if you want, or it's just affecting OC badly?

Could it be that the "deactivation" allows the threads to use the resources of the CU more efficiently (turning off a cluster may not turn off its resources, just the "thread" showing up as a core to windows) thus generating slightly more power usage?

I'm not sure how you mean it, but it definitely consumes less than all cores enabled (C2C):

IMG0033837.gif
From here.

[The Stilt]

Do you mean you cannot enable core 3, even if you want, or it's just affecting OC badly?

Not without core 2 being active (0-7) because core 3 is a "slave" core.

[chew*]

Heh against my better judgement I fired up rig one more time just to see if i missed something.

I found the wild card so to speak.

When disabling cores 2/4/6/8 ht must match NB.

When running 1/2/3/4 this appears to be a non issue.

It definitely needs further investigation.

Now the delta is not half as bad.

1/2/3/4 = 1.375 set for 4.8 gig

1/3/5/7 = 1.40 set for 4.7 gig

[Oese]

This was what i thought, there must be some quirks involved, physically there's no real cause for this imo..

[chew*]

This was what i thought, there must be some quirks involved, physically there's no real cause for this imo..

Well also you must keep in mind currently I am thrashing IMC purposely with mismatched 16g and 15gig usage in prime 95 as well + 64 bit win7 for a worst case scenario.

[Oese]

Aha alright

[Sgt.McRuff]

Chew what's the highest NB you gotten on air? And ln2?

[chew*]

Ok not quite what i was hoping for, have to test more benches....

So stability or a rough idea at least needed for 1/3/5/7, same identical settings will run 4.8 with 1/2/3/4.



Bottom line resource hit or not some apps are going to favor speed.......whether this only pertains to single threaded needs more testing. 3dmark 01 comes to mind.













Only the multi threaded bench gained with less clocks here.

[chew*]

Chew what's the highest NB you gotten on air? And ln2?

With 16g ram 2800 was best. Ln2 4500 ish through vantage.

[Manicdan]

i think i see what chew is getting at

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.

so i think the best thing is to just leave all cores there, OC high for full load, but then use turbo to OC lower threads even higher if possible.
so with a normal OCing air cooler, like ~4.7ghz on all CUs with a moderate voltage, but 5.0-5.2ghz for 1-2 CUs
i would like to know if using turbo with overclocking is as easy as it was with thuban, any insight on this guys?

[JPQY]

Hey chew* .. you also want to try this Chess test on your FX-4110 or/and on your other FX's you have!

You can easy select from 1 core to 32 cores ,ofcourse i want to put your best results in my list.
I use always chess engines to overclock my system and to run stable 24/7 .. just give it a try

You find explenation and software below in my signature.. setup need max. 3sec. if you have done the settings first time.

It can go even faster if you use Skype to help you with the setup.

JP.

[chew*]

Now I really have mixed feelings about running like this. It almost feels like a lose lose situation.

One of the key things you need to note here is not the overall scores or the cpu portions of the tests because it's quite obvious as was already observed without resource sharing multithreaded improves.......the issue is higher clocks benefit game tests even when resource sharing.

4600 results = 1/3/5/7
4800 results = 1/2/3/4