Maxxing CPU vs maxxing memory - results (AMD64 S939)

Jimbo Mahoney · 8 May 2006 at 21:36

OK, so I'm bored waiting for my 2GB RAM (two different sets) and extra 36GB Raptor to show up, so I thought I'd play with my max memory and max CPU clocks.

I've always been an avid "max CPU >> everything" when it comes to the S939 platform, but these results are interesting nonetheless, given the large difference between memory speeds and small difference in CPU speeds used.

I know my memory maxxes out @ 245 Mhz 2-2-2-6 1T
I know my CPU maxxes out @ approx. 3.05 Ghz

Given these two facts, we can come up with two possible configurations:

10 x 305, 140 divider = 3.05Ghz CPU + 217Mhz mem

and

10 x 294, 166 divider = 2.94Ghz + 245Mhz mem

That's 3.7% difference on the CPU and 12.9% difference on the mem.

Results:

Max CPU

Everest
Read = 6191
Write = 2559
Latency = 41.2

1M SuperPi
28.2

FEAR
47/107/278

3DMark '01
32244

Max memory

Everest
Read = 6768
Write = 2777
Latency = 38.2

1M SuperPi
28.68

FEAR
46/107/281

3DMark '01
31955

I was surprised that the max CPU cleaned up all round (except memory bandwidth, obviously). I expected SuperPi to prefer the latency and bandwidth offered by the max memory config.

I know a lot of you guys already know all this, but it's useful to have some more real-world results and to show the newer guys that CPU speed is where it's at on the S939 platform.

(If anyone's wondering why the 3DMark is so low, I had max details and LOD -4

see my sig for real 3DMark scores

)

Dutch Guy · 9 May 2006 at 09:58

With Socket 939 it is deffo CPU speed > everything else

I rate it in this order of importance of what should be maxed first: CPU speed > 1T > mem speed > mem timings

Although the mem timings/speed depends on what memorychips are used, some memorychips like high FSB with slower timings and other sticks like stock speed with tight timings.

What is slightly annoying is that memory benchmarks favor high memspeed and tight timings yet real world games do not mirror those results at all and the differences are much smaller than they are theoretical.

welshtom · 9 May 2006 at 10:53

Good work Jimbo, lets hope some of the memory guys come in here and see

Tom

Byron · 9 May 2006 at 11:49

nice1 jimbo

seems that mem bandwidth might be usefull to other software, audio/video/image rendering or editing though?

Minstadave · 9 May 2006 at 11:54

This is why AM2 isn't really going to do much for the AMD64 platform, plain old DDR provides plenty of memory bandwidth as it is, so DDR2 isn't really going to boost things.

Monstermunch · 9 May 2006 at 12:03

The main problem being that the AMD onboard ram controller can't handle over a certain speed, hence the reason why dropping the HTT from 1000mhz to 600mhz makes no difference in speed.

However I think you'll find that socket M2 is designed for the future and not just this year. I has been designed with quad core etc in mind. I also have no doubt that as things start to utilise dual core more effectively then the extra mem speed may be utilised.

Just some thoughts!

Minstadave · 9 May 2006 at 12:07

Monstermunch said:
The main problem being that the AMD onboard ram controller can't handle over a certain speed, hence the reason why dropping the HTT speed from 1000mhz to 600mhz makes no difference in speed.

I'm not convinced with that statement tbh. The reason upping the memory speed doesn't increase performance is the CPU is not memory bandwidth limited, so the CPU has plenty of bandwidth for the work its doing.

The memory bandwidth isn't linked to the HT bus at all going by this document, so dropping the HT speed should make difference to memory performance:

http://download.microsoft.com/download/1/4/7/1470554c-eaa8-4e58-a352-972b9b70087f/AMD.ppt

The reason dropping the hypertransport (not HTT) speed from 1000mhz to 600mhz makes no difference is that the hypertransport bus is overengineered and provides plenty of bandwidth from the CPU to the outside world.

Dutch Guy · 9 May 2006 at 12:15

Minstadave said:
This is why AM2 isn't really going to do much for the AMD64 platform, plain old DDR provides plenty of memory bandwidth as it is, so DDR2 isn't really going to boost things.

AMD is going DDR2 because that is where most of the development is but other than that there is no advantage (yet)

I think I read in an article comparing different DDR2 timings that it is the same story and that for the most part timings also make little difference with DDR2