Memory performance
Obviously, the biggest difference between these chipsets is memory support, so let's cut right to the chase. We'll start with SiSoft Sandra's memory tests, which measure bandwidth using a modified, more intensive version of Stream. These tests are a measure of memory performance only—and the Sandra tests, of bandwidth in particular—so don't get too hung up on them before you see results from other benchmarks.

It's no surprise that the 850 chipset with dual RDRAM channels comes out on top here, but there are a few surprises beyond that. First, the P4X266 does especially well, pushing around about 1100MB per second. That means DDR SDRAM is realizing over half of its bandwidth potential here, while the RDRAM system comes up shy of 50% of its 3.2GB/s theoretical peak.

Next, the PC133-based 845 system actually delivers more bandwidth in Sandra's integer test than the DDR-equipped Athlon system. That situation changes in the floating-point test, but obviously the combination of the Pentium 4 and 845 chipset is able to make the most of PC133's available bandwidth.

Finally, the Pentium III system, also based on PC133 SDRAM, doesn't deliver nearly as much throughput as the 845. Although the Pentium III "Tualatin" processor is equipped with hardware prefetch logic, it just can't keep up with the P4 here. No doubt the PIII's 133MHz front-side bus is a handicap.

Now let's look at these same numbers in a little bit different way using Linpack.

Linpack graphs aren't the easiest to read. Take a second to look at the axis labels, though, and you'll get it. We're measuring processing throughput here, in megaflops, for data matrices of different sizes. The small matrices, on the left half of the graph, fit into the processors' L1 and L2 caches, so processing throughput is high. On the right half of the graph, where the matrices are too large to fit into the caches, performance drops. It's there, at the larger data sizes, where we get a better sense of main system memory bandwidth. Overall, the shape of the graph gives us a nice visual picture of how a system's tiered memory architecture performs.

Concentrate on the Pentium 4 systems, and you'll see identical throughput up to about 192K. Once we fully leave the confines of the P4's 256K L2 cache, the lines for the three chipsets diverge. The RDRAM-based 850 delivers the most sustained bandwidth, but the DDR-driven P4X266 isn't far behind. The numbers for the PC133-equipped 845 system are much lower, but they still overlap with the numbers for the Athlon DDR rig. All in all, these results are fairly similar to the Sandra results above.

Bandwidth is only one component of memory performance, however, and not always the most important one. There's also latency, or the lag time involved in memory accesses, to consider. Here's how things stack up in terms of latency:

Now we can see how the Pentium 4 chipsets stack up in a different sort of memory performance, and it's a very different picture. The 850 chipset with RDRAM comes in last, well behind the P4X266, and even behind the bandwidth-starved 845. Latency is the Achilles' heel of RDRAM, and it's much of the reason Rambus-based systems don't always deliver the real-world performance their theoretical bandwidth peaks seem to promise. RDRAM's high clock speeds and narrow, 16-bit interfaces deliver lots of data, but lag times are a little higher than they are for SDRAM.

Surprisingly, both of our non-P4 comparison systems turn in much lower latency numbers, though the Pentium 4 beats them soundly in bandwidth tests. In general terms, latency and bandwidth are interrelated; low latencies lead to high bandwidth and vice-versa. (Think of ping times on a 1.5Mbps connection versus a 33Kbps connection, and you're getting the idea. RDRAM is a notable exception, of course.) In this case, however, it's possible the Pentium 4's higher memory latencies are related to its high memory bandwidth scores. The Pentium 4's liberal use of hardware prefetch logic may be saturating the memory bus, driving up memory latencies.

Maybe.

Whatever the case, memory performance is only part of the overall performance picture. For our purposes, it's most important to note that the P4X266 has a particularly nice combination of memory bandwidth and access latencies. The 845 is slower both in terms of bandwidth and latency.