Is 3:2 For You?

Quibbles

Overclocking Experience

The CPU is a 2.4C, a week 11 Malaysian CPU, watercooled.

No problems with installation outside of getting SATA to work right.

First overclock was at 3GHz, 250MHz FSB, 200MHz memory speed. No problem at default CPU voltage. Set the RAM at 2.7V since even Corsair says that’s OK, no problem running 2-2-6-2 (or even 2-2-5-2) speed.

Second overclock was at 3.3GHz, 275MHz FSB, 220MHz memory speed. CPU voltage had to be upped to 1.6V, memory had to be slowed down to 2.5-7-3-3, but that was fine.

Further attempted overclocks began blowing up. Best I could do with very slack memory setting at 2.8V was 280MHz FSB/224MHz memory, with questionable stability.

Decided to eat my own dog food at this point and began running at a 3:2 ratio. Maximum stable point was at 295/196, or 3.54GHz at 1.675V. Prime95 blows up right away at 300MHz, it blows up fast at 295MHz at 1.65V.

That’s not the interesting point. What happened with memory speeds while I was doing this was.

Ed Stroligo

Conventional Wisdom

Up to now, conventional wisdom has stated that one should maximum memory speed at practically any cost. This is supposed to increase bandwidth and thus speed up your computer.

Well, from what I’ve seen, conventional wisdom needs some serious modification.

Look at these memory benchmarks.

First, 1:1 ratio: FSB at 200MHz, memory at 200MHz. Memory settings at 2-2-6-2:

Next, 5:4 ratio: FSB at 250MHz, memory at 200MHz. Same memory settings. Please note, the memory is running at the same speed it did before:

Hmmmm. Everything’s absolutely the same except the FSB, but the memory score goes up about 18%.

This time, FSB at 295MHz, memory at 196MHz. Same memory settings. Please note, the memory is actually running a little slower than the first two times.

Hmmmm. We slow down the memory and it goes faster.

How does it do running quickly?

Ed

220MHz FSB, 220MHz memory. Memory settings 2.5-7-3-3. That’s just about the best I can do with the Corsair (and the slower memory setting drops the Sandra score about 160 points for both measurements):

Faster than 200/200, but slower than 250/200 and 295/196.

Let’s try 275MHz FSB, 220MHz memory. Memory settings 2.5-7-3-3:

A bit faster in measurement one, a bit slower in measurement two, despite running 12% faster. The slower memory settings drop the 220MHz figures by about 3%, so even if you add 3% to the 220MHz figures; it still comes nowhere near what you would normally expect from the extra speed.

Just to summarize the basic numbers:

200/200: 4603/4593
220/220: 4954/4977
250/200: 5402/5416
275/220: 5881/5761
295/196: 5814/5807

If you start removing buffering and use of SSE2 when benchmarking with Sandra, most of the phenomenon goes away, but it’s still there.

Another Measurement

What does PCMark 2002 report for memory scores?

Here’s just the raw scores:

200/200: 8118
220/220: 8667
250/200: 9206
275/220: 9824
295/196: 10479

The effect is even more pronounced than in Sandra.

Ed

How Could This Be?

What we may be seeing here is the relieving of a bottleneck we didn’t know was there, much like the FSB/memory with the Athlon. Remember that two sticks of RAM running at 200MHz is theoretically capable of 6400MB/s.

Increasing the FSB increases the maximum theoretical bandwidth of the system, from 6400Mb/sec at 200MHz to 8000Mb/sec at 250MHz to 9600Mb/sec at 300MHz. These scores would seem to indicate that increasing FSB gives memory more of a chance to operate at maximum efficiency.

Does This Have Any Real World Impact?

Don’t know yet.

However, there’s only two possibilities: it does or it doesn’t.

If it does, people may find that by running 3:2, they can run (perhaps cheaper) memory at reasonable speeds (and voltages) and hardly lose anything by it. Indeed, if their overclocking is being held back by memory, they may well find that they can overclock the processor a bit more and end up with a net gain.

If it doesn’t, then these memory benchmarks are pretty worthless with the new Intel chipsets.

Ed