CPU Speed vs. Memory Bandwidth

[Insomniak]

So, 2.4Ghz is stable (see sig) but I have to run my memory clocked down slightly (DDR370 vs DDR400).

Memory bandwidth benches are as follows:

SANDRA
DDR400: 5563/5502
DDR370: 5269/5271

Everest
DDR400: 5241 Read/2304 Write
DDR370: 5119 Read/2150 Write


So apparently running this overclock and downclocking the memory results in a negligable loss in memory bandwidth. I have a few subjects here:

1) These are synthetic benchmarks. Anyone got a better idea of how to tell what the real world effect will be?

2) Do you guys (and gals?) think it would be more prudent to run at 3500+ speeds with full DDR400, or keep the current OC of 3800+ speeds and leave the memory slightly downclocked.

In regard to #2, being that I'm running Dual Channel memory anyway, I think I've got more than enough memory bandwidth to spare, even at lower speeds. I'm leaning towards keeping the higher CPU clock.


Edit: Forgot to mention - the primary purpose of this rig is gaming.

 

[DAPUNISHER]

The formula is simple for A64, clockspeed > * Zebo's thread should give you a good idea about how few areas the timing and bandwidth benefit link

The memory controller comes along for the ride when you overclock the A64 which is partially why performance scales so well with clockspeed on A64.

 

[Insomniak]

Originally posted by: DAPUNISHER
The formula is simple for A64, clockspeed > * Zebo's thread should give you a good idea about how few areas the timing and bandwidth benefit link

The memory controller comes along for the ride when you overclock the A64 which is partially why performance scales so well with clockspeed on A64.

I thought that was probably the case - as you mentioned, with the memory controller on-die, I figured external clocks and timings wouldn't matter much. Everything doing the work is getting its clockspeed jacked along with the core.

Also tightened the timings up a bit. Since the RAM is running underclocked, it certainly shouldn't mind, and memtest says it's perfectly happy. Regained around 50 - 80% of the lost bandwidth by doing so, according to SANDRA/Everest. For comparison:

SANDRA
-------------DDR400: 5563/5502
-------------DDR370: 5269/5271
DDR 370 @ 2-3-3-7: 5313/5446

Everest
-------------DDR400: 5241 Read/2304 Write
-------------DDR370: 5119 Read/2150 Write
DDR 370 @ 2-3-3-7: 5218 Read/2205 Write


Again, real world, this probably doesn't mean much, being that single channel and dual channel show minimal differences in performance, despite the wide gap in bandwidth. However, it's always nice to have your system humming along at full capability.

 

[sangyup81]

If this rig is for gaming, don't compare Sandra and Everest scores.

It's real easy to do a HL2 Stress Test or download 3dmark.