VIA's KT333: The Athlon gets DDR333
by Anand Lal Shimpi on February 20, 2002 1:57 AM EST- Posted in
- CPUs
Memory Bandwidth
As we move into the new year we start encountering the 2002 versions of many popular benchmarks. One such benchmark that has actually been out for a while now is SiSoft Sandra 2002. The one thing we have used Sandra for in the past is as a useful measure of memory subsystem performance, in particular its memory bandwidth test is usually fairly accurate.
In the 2002 version of Sandra, the benchmark now includes the ability to perform the conventional memory bandwidth test but also enables using SSE instructions to move the test data around. The end result is that you can get a much better idea of what the true peak theoretical bandwidth of a particular platform happens to be. We performed both the conventional memory bandwidth tests with the enhanced instructions disabled and with them enabled.
|
1986Using the enhanced data movement instructions it's clear that we run into the 133MHz FSB limitation of all of the platforms. The 2% difference in memory bandwidth between our KT333 and KT266A test platforms is small enough to be attributed to variations in board designs. Interestingly enough is that although the nForce 420-D platform has twice the theoretical bandwidth of the KT266A, its attainable memory bandwidth is actually around 8% lower. This could be because of NVIDIA's implemention of the EV6 FSB interface or because of their memory controller in general. The reason this doesn't translate into an 8% reduction in real world performance is because there are never any situations where you'll hit this sort of memory bus utilization.
This perfectly illustrates the first point we made about DDR333's usefulness in today's platforms; you shouldn't expect to see a tangible performance improvement because in the end you'll be limited by the Athlon's FSB.
|
Disabling the enhanced instructions simulates a much more down-to-earth usage pattern that is still a bit on the high-side for the memory bandwidth utilization characteristics of most applications. Here we can see that although the peak theoretical memory bandwidth figures don't increase by much, the actual attainable bandwidth figures do go up a bit.
The reasoning behind this is that, all things kept the same, as clock speed increases latency naturally decreases. Lower latency memory accesses results in greater memory bandwidth utilization and thus the increased figures you're seeing out of DDR333 SDRAM used with the KT333 vs. DDR266 SDRAM used on the KT266A.
What's even more interesting is that the nForce's dual channel memory controller setup on the 420-D is able to provide an additional 7% of bandwidth over the single channel setup.
As we've seen numerous times in the past however, you cannot base your real-world performance expectations on synthetic tests like these because in the real world, the ideal scenario is never the case.
From this point on we've omitted the nForce 220-D results since they are identical to the nForce 420-D results in all real-world performance tests where the integrated graphics are not used.
Facebook
Twitter
RSS
0 Comments
View All Comments