E8400 Synchronous vs. Asynchronous?

[Somniferum]

So I finally got around to overclocking my E8400 (E0). It runs great @ 3.7GHz, slightly undervolted to 1.2v (BIOS). Temps are around 38c idle, 54c full load (Orthos small FFTs).

My question is about the RAM. I'm running 4Gb of OCZ DDR-800 in dual channel. (And I can finally see all 4Gb now that I'm on Win7 x64, yay.) By default it runs with a 5:6 divider, which puts it at 988Mhz currently (494 x 2). But I keep seeing references to some magical benefit to running in synchronous mode instead, especially with C2Ds it seems.

So would it somehow benefit me to set the RAM to 1:1 and run it stock @ 800MHz (technically 824, but close enough)? It seems counter-intuitive to me that lowering the speed from 988 to 800 would be a good thing, especially with the CPU @ 3.7 and hungry for bandwidth. But since I don't really understand what the great benefit of synchronous mode consists of (despite searching) I can't really be sure how to evaluate the difference. Hence this post.

I'm not really concerned about RAM timings -- this is cheap stuff and I tend to just leave it at SPD. So that's not a factor to consider.

So, should I run it at 988 asynchronous, or 800 synchronous? Or some other setting I haven't even considered (like lowering the multiplier)?

Any thoughts appreciated.

 

[Spoelie]

In sync should provide the lowest latency and best performance (since the fsb is saturated) in an ideal world, however:
* core2 is not a very memory-latency sensitive architecture
* the cpu is not the only consumer of memory bandwidth, DMA devices (even GPUs) can bypass the cpu and access memory concurrently - increasing memory bandwidth demand

The only way to find out for sure is to run benchmarks... The various 3Dmarks might show differences.
Truth be told, if there is any difference, regardless of which one is better, it will be very slim

 

[Somniferum]

Originally posted by: Spoelie
In sync should provide the lowest latency and best performance (since the fsb is saturated) in an ideal world, however:
* core2 is not a very memory-latency sensitive architecture
* the cpu is not the only consumer of memory bandwidth, DMA devices (even GPUs) can bypass the cpu and access memory concurrently - increasing memory bandwidth demand

The only way to find out for sure is to run benchmarks... The various 3Dmarks might show differences.
Truth be told, if there is any difference, regardless of which one is better, it will be very slim

Thanks ... I figured the difference would be slim, but it's the principle of the thing as well ... I built this system and it gives me satisfaction to know it's configured optimally. Guess I need to run some benchmarks. Super Pi seems to take 13s for 1M regardless of what I do to the RAM, so I guess I need to break out Sandra or something that measures actual memory bandwidth and latency.

To further complicate things, I decided to research the memory timings since that seems to play into this issue. Turns out this particular OCZ RAM is kind of ... strange. Supposedly it's rated at 5-4-4-15, but that's only if you overvolt it to 2.1v (which is out of spec for this platform). Since I've been running it stock @ 1.8v, it's been defaulting to 6-6-6-18, which can't be good. So I guess I need to pump up the volts, fix the timings, and see if it will even still run @ 988MHz.

 

[Somniferum]

I thought I should provide a conclusion in case anyone finds this thread in the future. The benchmark scores are in, and asynchronous wins. I had to loosen the timings a bit to 5-5-5-15 (remaining at 2.1v), but at 5:6 the memory bandwidth clocked in at 8Gb/sec compared to 7.7Gb/sec at 1:1. So as usual, clock speed trumps timings. No big surprise there, it's what I intuitively suspected all along. But I kept seeing all these references to the holy grail of 1:1 synchronous mode with C2D. Just goes to show how wrong the conventional wisdom can be.

 

[n7]

There's no holy grail of 1:1.

It's a a many-years-old very inaccurate myth with Intel platforms.

Only exception is nForce chipsets.

Glad you tested for yourself :thumbsup:

 

[amenx]

Originally posted by: Somniferum
I thought I should provide a conclusion in case anyone finds this thread in the future. The benchmark scores are in, and asynchronous wins. I had to loosen the timings a bit to 5-5-5-15 (remaining at 2.1v), but at 5:6 the memory bandwidth clocked in at 8Gb/sec compared to 7.7Gb/sec at 1:1. So as usual, clock speed trumps timings. No big surprise there, it's what I intuitively suspected all along. But I kept seeing all these references to the holy grail of 1:1 synchronous mode with C2D. Just goes to show how wrong the conventional wisdom can be.

Yes, I also came to that conclusion long ago, asynchronous wins when the ram mhz is higher.

 

[Gillbot]

I've always gotten the best performance when running 1:1 but I'm usually running my C2D chips in the 500FSB range anyway so the clock speed is high anyway.

 

[n7]

Originally posted by: Gillbot
I've always gotten the best performance when running 1:1 but I'm usually running my C2D chips in the 500FSB range anyway so the clock speed is high anyway.

If you could get your RAM higher than the 500 MHz range, like say 5:6 (DDR2-1200), it'd be faster

Obviously in your case, the limitation is the RAM, so 1:1 is the only choice.

But when there's the ability to run the RAM at a higher speed, i'd definitely do it.

 

[Gillbot]

Originally posted by: n7

Originally posted by: Gillbot
I've always gotten the best performance when running 1:1 but I'm usually running my C2D chips in the 500FSB range anyway so the clock speed is high anyway.

If you could get your RAM higher than the 500 MHz range, like say 5:6 (DDR2-1200), it'd be faster

Obviously in your case, the limitation is the RAM, so 1:1 is the only choice.

But when there's the ability to run the RAM at a higher speed, i'd definitely do it.

I have, the gains were minimal if any at all and the stress on the ram wasn't worth it.