[how to] Get the best SuperPi efficiency out of the K10 platform

[Zeus]

The aim:
As the K10 platform has matured more people start benching their Phenom II cpu's in all known bechmarks.
As SuperPi has become a standard here more or less it's good to know how to get the best result out of your K10 system knowing this platform isn't as strong as others in SuperPi.
I tried a lot of relevant settings to see what works best and what not, in other words, use it to your advantage.

The method:
The effficiency of the system is gauged by the so called Performance Product (PP) and is basically the CPU speed in MHz x the output result of SuperPi.

The hardware:
Motherboard- Asus M4A79 deluxe bios 0903
CPU- Phenom II 940 black edition running on all 4 cores
RAM- Cellshock Black PC6400 2x1GB

The sofware:
Winows XP Professional SP2, no tweaks, just a clean install.

The testresults:
Let's start off pretty basic, i tried to have all the results at around 3.6GHz or at least close to that as possible.

Here is 18x200 RAM at DDR1070 with 5-5-5-15 2t timings, NorthBridge at 2GHz:



PP=3612x19.203=69361 This should be easy to improve.
Let's just raise the NB to 2.6GHz :



PP=3612x18,906=68288. That's a lot better, only by just raising the NB frequency.

Now let's see what tighter timings will do.
18x200 RAM at DDR 800 4-4-4 2t NB at 2.6GHz:



As you can see, exactly the same as with DDR 1070 5-5-5.


PP=3612x18,906=68288

Now using the 1:2 divider gives the advantage of using tighter timings and also allows 1t timing, i was really curious how much 1t would bring on this platform.

18x200 RAM at DDR 800 4-4-4 1t NB at 2.6GHz:



PP=3612x18.922=68346
WTF? It's worse than 2t! I tried several times but this was the best i could get.
To the best of my knowledge command rate is the timing of the bus between the cpu and the memory, at 1t the bus gets data at every clockcycle, at 2t every other clockcycle.
Appearently due to the integrated memory controller the latency is that low that 1t makes zero impact on the performance or makes it even worse.

So, basically forget 1t for the K10 platform if you're planning to run SuperPi, DDR3 might differ here.

So what else do we have? Well even 4-3-3 1t timings were possible at the 1:2 divider so why not?

18x200 RAM at DDR 800 4-3-3 1t NB at 2.6GHz:



PP=3612x18,906=68288

No real improvement here, now let's raise the NB frequency.

18x200 RAM at DDR1070 with 5-5-5-15 2t timings, NorthBridge at 2.8GHz:



PP=3612x18.828=68006. Now that's better! Still not sub 68K but we're getting close now.

How about raising the busspeed?

15x240 RAM at DDR 800 4-4-4 1t NB at 2.6GHz



PP=3600x18.985=68346. No, that's not the way. We could use higher busspeed to our advantage cause with raising the busspeed the memoryspeed goes up as well.

Now 14x260 RAM at DDR 866 4-4-4 NB at 2.6GHz



PP= 3640x18.812=68475. Not really what i was looking for but maybe some more speed on the RAM?

14x260 RAM at DDR 1040 5-5-5 NB at 2.6GHz



PP=3640x18.765=68304. Getting better now. Unfortunately the board refused to boot or run stable above 260 bus so i couldn't get more RAM speed, i'm sure more speed on the RAM would have improved results.

The last thing i could try was running more NB speed.

Here is 14x260 RAM at DDR 1040 5-5-5 NB at 2.86GHz:



PP=3640x18.688=68024 Still no sub 68K but there's one final trick up my sleeve...

14x260 RAM at DDR 1040 5-4-4 NB at 2.86GHz:



PP=3640x18.671=67962. So there you have it, sub 68K

For those who think Windows7 is faster:



PP=3612x19.031=68739.

I'd rather stick to XP.

I hope you guys enjoyed this little write up.

[haylui]

thanks for the in-depth research in making K10 run faster on SuperPI

[Particle]

Thanks for the work in investigating this. Perhaps a chart would make the presentation & comparison between the various settings easier for the viewers.

[Quintero]

WTF? It's worse than 2t! I tried several times but this was the best i could get.
To the best of my knowledge command rate is the timing of the bus between the cpu and the memory, at 1t the bus gets data at every clockcycle, at 2t every other clockcycle.
Appearently due to the integrated memory controller the latency is that low that 1t makes zero impact on the performance or makes it even worse.

So, basically forget 1t for the K10 platform if you're planning to run SuperPi, DDR3 might differ here.

Command rate has no direct influence on the databus or data transfer rate. What it changes is the number of clock cycles the controller uses to qualify new commands and their corresponding addresses. As you add more memory sticks, the capacitive load on the controller's address unit increases, which in turn reduces the slew rate of the signals. Since the signals switch slower from 0 to 1 (and vice versa), the controller might need an additional clock cycle to generate the correct combination of 0s and 1s. The biggest drawback of 2T is probably not the delay itself, but rather that the address bus gets more "crowded". Each entry takes up two timeslots, so there would simply be less space (time) to squeeze in new commands.

The difference was rather big in the 939 days, but small for AM2 (although 1T was still slightly faster).

In K10's unganged mode, the difference should be even smaller, because there are two independent controllers, which gives potentially longer read/write bursts per command. This lowers the risk of collisions on the address bus. A transfer of a given data chunk would simply require fewer commands (per controller), so fewer entries would have to "compete" for the same timeslots.

As a result of the "new" Fly-By address bus topology and the on-DIMM termination, DDR3 systems can handle 1T at reasonably high loads and clockrates.


Anyway, nice work! I enjoyed analyzing the results

[LowRun]

Nice work there

[Zeus]

Thanks for the work in investigating this. Perhaps a chart would make the presentation & comparison between the various settings easier for the viewers.

I might just do that, thanks.

Quintero,

Thanks for shedding a light onto the command rate part, very interesting and much appreciated.

[knopflerbruce]

This is sweet! Too bad you didn't get 520 4-4-4 That would get your time down a bit, I think. What about ganged vs unganged, btw?

[Zeus]

Thanks. I would have loved to run 520 4-4-4 but the 3:8 divider won't let me boot at cas level 4.

Ganged vs unganged never crossed my mind to be honest, i might as well give that a try later, this testing is very time consuming stuff you know.

[chew*]

I'm working on this atm, trying to get this gigabyte to cooperate at 300fsb....

Here we go, I would not reccomend this to anyone for everyday use, this particular board is a PITA to get 300 fsb stable on and the voltage I am feeding everything is , I could probably get the efficiency better with higher ram speed but there is fsb wall at 300+, I will try a diff configuration later...

The times don't look good but the efficiency is just over 67K,

[techtrancer]

@ Zeus

realy good job! thank you!


@ chew

i have the same problem with my asus MVP

if i would make 301 in the bios the mainboard wants to boot an CPU Clock auf 4200mhz with every CPu multiply

but if i boot with 300 and pust it in windows with AOD over 300 it is no problem

[chew*]

yah tried aod, even with max voltage everywhere on the board it doesn't want to here it.

[roofsniper]

nice, very informative. a little off topic but has anyone found a time where tightening the timings and changing to 1T has actually had a benefit to performance? i have changed my timings all around and never once have i seen a difference in performance.

[techtrancer]

@ chew*

did you have Micron D9GMH to make a test with them?

maybe your 2 x 2gb memmory does not work with more than 600

[chew*]

I tried a 2x2g set of corsair dominators v1.1 and it puked them up, hosed my OS twice, oddly they passed memtest fine.

Had to run 5-4-4-10 with the 4gig kit though........didn't want to hear CL 4.

Someone donated me a benching set 2x1g dominator v1.2, I will give them a shot.....

My board and mem can clock decent, no clue what these sticks are but they crank at rated timings.....they are preety much dead now though.

[roofsniper]

my corsair dominators always passed mem test but they would crash in prime and had to be overvolted to run stock clocks and timings.

[techtrancer]

@ chew*

forget memtest for ramtest
did you know S&M ?
it is the best ram test i ever use! much better than memtest ore prime

here is the DLL http://rapidshare.com/files/198699292/S_M.exe

here are the best settings for the ram test



if you search for e good benchmark DDR2 ram, you can have some from me. but i am in austria in europ!

[Slagathor]

to bad S&M doesnt work for 64bit......

[G0ldBr1ck]

I have 2 sets of 2x2 dominators V1.1 and V1.2 and they both do 4.4.4.10 at 1000Mhz. I have faith that those 2x1g set will do the same easy Chew*. Sadly one of my 2x2 sets died so I RMA'd them and got back a set of V4.2 And they wont do any better than 5.5.5.12. I have some Old (at least a year) Gskill ram 2x2 im testing right now at CAS 4, gonna start memtest and go to bed.

[techtrancer]

to bad S&M doesnt work for 64bit......

i use XP32 and XP64 and there it works. oese use Vista 64 and there it also runs.

you must unpack it first

[massman]

It's all in the CPU/NB frequency balance: if the NB is clocked a lot lower than the cpu, your efficiency will drastically decrease (performance product will increase)

[Zeus]

I'm working on this atm, trying to get this gigabyte to cooperate at 300fsb....

Here we go, I would not reccomend this to anyone for everyday use, this particular board is a PITA to get 300 fsb stable on and the voltage I am feeding everything is , I could probably get the efficiency better with higher ram speed but there is fsb wall at 300+, I will try a diff configuration later...

The times don't look good but the efficiency is just over 67K,

Would you like to share which voltages you raised to get 300fsb?

Your efficiency is good due to high ram speed and high NB clock not because of 300 fsb but i think you know that.

Keep them coming and show me sub 68K chew*!

@ Zeus

realy good job! thank you!


@ chew

i have the same problem with my asus MVP

if i would make 301 in the bios the mainboard wants to boot an CPU Clock auf 4200mhz with every CPu multiply

but if i boot with 300 and pust it in windows with AOD over 300 it is no problem

Thanks. Very nice efficiency there, 600+ on the ram seems to do a great job, i might give that a try too.

It's all in the CPU/NB frequency balance: if the NB is clocked a lot lower than the cpu, your efficiency will drastically
decrease (performance product will increase)

Very true my friend, NB clock seems to have the most impact on the score with the memoryspeed in second place. Timings seemed to matter a lot less.

[techtrancer]

i think much memmory is also good for superpi!

4gb are better than 2 gb

[chew*]

Would you like to share which voltages you raised to get 300fsb?

Your efficiency is good due to high ram speed and high NB clock not because of 300 fsb but i think you know that.

Keep them coming and show me sub 68K chew*!

Zeus If you want it via PM no problem however I'd like to refrain from posting them for all to see, not for the faint of heart and people will be blowing up cpu's/mobos left and right.

[massman]

Currently working on the second part of my Phenom performance scaling analysis (first part was about the Ph-I) and i just hit 67947 - no tweaks whatsoever.

Will upload the screenie later

//edit: 67763

[massman]

67579, this is easy. I added task manager.



I must have a good chip after all: 2.8GHz NB without touching the voltages (except from the cpu, just to be safe when testing)