DDR: What Does What

You’ve seen DDR memory scores all over the place. You don’t know what to make of them.

There are three factors which influence memory scores:

1. FSB Speed
2. Memory settings
3. Overall CPU speed

The first two are more important than the third.

FSB Speed

Here’s two sets of SiSoft memory benchmarks run at roughly the same system speed, using an MSI K7Master and two 256Mb Crucial PC2100 modules, using beta BIOS 1.12 and Win2K SP1 (Win98 yields scores within 1-2% of Win2K). The first set is run at “slow”
memory settings, the other at “fast” memory settings.

10 X 133Mhz FSB, Slow Memory Settings

9 X 149Mhz FSB, Slow Memory Settings

We increased the bus speed about 12%, and got improvements of 8% and 12% in the two scores respectively.

Let’s give another example:

10 X 133Mhz FSB, Fast Memory Settings

9 X 149Mhz FSB, Fast Memory Settings

Again, we increased the bus speed 12%, got a 9% and 11% increase in the scores.

Memory Settings

More aggressive memory settings give you a higher memory (provided it works). But what are aggressive memory timings?

Most know about setting RAM to 4-way interleave. That appears to built into at least the MSI BIOS, for Via systems,
you can either change that setting in the BIOS or by using WPCREDIT.

The following are the “optimized” memory settings for the K7Master:

SDRAM PH Limit: 8 cycles
SDRAM Idle Limit: 8 cycles
SDRAM Trc Timing Value: 8 cycles
SDRAM Trp Timing Value: 3 cycles
SDRAM Tras Timing Value: 7 cycles
SDRAM CAS Latency: 2 cycles
SDRAM Trcd Value: 3 cycles

These are the “slow” settings. This is what I changed them to:

SDRAM PH Limit: 8 cycles
SDRAM Idle Limit: 8 cycles
SDRAM Trc Timing Value: 7 cycles
SDRAM Trp Timing Value: 2 cycles
SDRAM Tras Timing Value: 5 cycles
SDRAM CAS Latency: 2 cycles
SDRAM Trcd Value: 2 cycles

And here are the results:

10 X 133Mhz FSB, Slow Memory Settings

10 X 133Mhz FSB, Fast Memory Settings

By changing the settings, we got a 9 and 10.5% increase in the benchmarks.

Another example:

10 X 149Mhz FSB, Slow Memory Settings

10 X 149Mhz FSB, Fast Memory Settings

Here, we get a 9% and 9.5% increase in the scores.

CPU Speed

If you just increase the CPU speed, you don’t get as much improvement.

Some examples:

8.5 X 133Mhz FSB, Slow Memory Settings

10 X 133Mhz FSB, Slow Memory Settings

We increased the CPU speed 18%, but only got a 4% and 3.5% increase in the scores.

Another example:

9 X 149Mhz FSB, Fast Memory Settings

10 X 149Mhz FSB, Fast Memory Settings

We increased the CPU speed 11%, but only got a 3% and 2% increase in the scores.

Some comments about Sandra:

I found a bit of variance in the scores. It tended to improve slightly on multiple runs by 1-2%, once in a blue moon by even more.

Unlike most other memory tests, this one was conducted with 512Mb of RAM, but this didn’t seem to have any significant bearing on scores.

However, I found out that Sandra’s default only tests half of the RAM used. I tried to change the option to test all the RAM, but the program hung up. This
was not an indicator of stability; the program hung up at all speeds and settings.

How Do These Scores Look Compared To Other Websites?

With “fast” memory settings, I was getting just about the same scores reported by Tom’s Hardware. Tom’s Hardware did report the memory settings used, but most other websites don’t.

(I suspect the MSI Via board tested could have been tweaked a bit more using WPCREDIT, but it would have made only a couple percent difference).

This is going to be important in product comparisons. As you can see, just a few changes in memory setting can make about a 10% difference in scores. If you don’t know what the settings were on the platforms being compared, the “winner” may be the winner just due to memory settings.

In the case of at least the K7Master, the memory settings are tweakable, but 4-way interleaving is nowhere to be found. Given the scores, I presume it’s turned on, but I can’t swear to it.

For at least the initial Via chipsets, this can be a more formidable task. MSI has been tweaking the BIOS, as well as some individuals using WPCREDIT (changing the setting in position 64 turns on 4-way interleave). If you go with the latest BIOS and tweaks, you have
what looks to be a more-or-less even playing field with the AMD chipset.

Who wins?

The impression I’ve gotten is that the K7Master will beat the K7T Pro266 in memory scores by about 6-7% on board where the resistor is at R127. If the resistor is at R126, the Via board draws even, or maybe 1-2% better.

Is this meaningful in the real world? I greatly doubt it. I’m certainly going to test this out in the upcoming days, but that 6-7% memory edge will probably yield a 0-1.5% improvement. It will be closer to 0% in business apps, closer to 2% in memory-sensitive games.

However, memory scores aren’t everything. At the moment, people seem to be having more problems with the MSI Via than the MSI AMD board (though I’m by no means ready to give the K7Master a clean-bill-of-health yet; I’ve had a few difficulties so far).

Memory benchmarks don’t mean a lot to overall performance. However, many find it important, but unless you know just what the memory settings were in any product comparison

Conclusions

Your emphasis should be on FSB and memory settings, not CPU speed, when memory speed is important to what you run. Try to get those two up as high as possible consistent with stability.

If it’s a choice between a higher FSB or more aggressive memory settings, experiment; it will probably end up six is a half-dozen of the other. Theoretically, overclocking FSB will yield improvement elsewhere, so try to go with that.

Experiment. The measurements I’ve given here are meant to be a tool, not a rule. If your machine behaves differently, go with what your machine like, not me.

Email Ed