Iwill DP400 (Cont'd)

Clibench Mk III SMP 0.7.15

Clibench is another SMP aware synthetic benchmark that we're quite fond of around here. There has been some speculation on its validity as a test for newer SMP platforms, but if we compare the Clibench numbers to the Sandra numbers above, it paints a pretty accurate picture. So, at least for the time being, we'll stick with it.

It is interesting to watch the E7500 and i860 platforms trade back and forth in the Clibench tests. As you can see though, memory bandwidth isn't everything. The E7500 beats the i860 in almost all tests, so maybe there is hope for "Placer". The Athlons put up a strong showing in the FPU intensive tests and beat both Xeon platforms in almost every test. It's a pretty safe bet to say that Clibench is not SSE2 optimized.

Cachemem 2.6

Since one of the main "features" of the i860 chpset is its use of RDRAM it seemed fitting to add a few more benchmarks to our suite. When Intel announced that they would only be supporting RDRAM people were up in arms about the cost of the memory and its added latency. Intel has since started to back away from RDRAM, but is that a good thing? I had also been hearing rumors about i860 boards losing memory throughput when multiple (more than two) sticks were installed at the same time. Taking all this into account, I tracked down Cachemem, Linpack and a bunch of different sizes of RIMMs and headed down the road to RDRAM enlightenment.

First up is a handy little memory benchmark by the name of Cachemem. Cachemem runs from a commandline and analyzes cache and main memory performance for reads, writes and latency using a variety of "block" and "step" sizes. Basically it sends a chunk of information through the cache and system memory and records the time taken, rinse, repeat with different sized blocks. It gives incremental numbers as well as one final number for "overall" performance.

I was curious to see how the number of installed RIMMs hurt performance, or if it even hurt anything, so I ran the tests with two 256mb modules installed with two CRIMMs, as well as with four 128mb modules and no CRIMMs. You'll also notice that I didn't include the E7500/DDR combo in these tests. These benchmarks are new to us, so until I review my next E7500 board (I have one here now), it'll remain a mystery. Another big questionmark is how the i860 boards with MECs (and consequently, another MCH) will handle the same tests. My next i860 board review will be using a MEC, so that too will soon be answered. But anyway... Back to the task at hand.

In the overall performance numbers I saw something that was completely unexpected, at least to me. The four 128mb RIMMs actually outperfom the two 256mb modules. Granted, the difference isn't huge, but it is there. The only thing I can think of is that the CRIMMs add a bit of latency which hurts throughput a bit. Since speculation is for losers, let's find out for sure!

For the latency tests I settled on a block size of 2048bytes, with a 512k step. It is a fair midpoint in the tests, and I have seen other sites use the same block and step sizes, so our numbers should compare well to the ones you might see elsewhere.

It looks like my assumptions were correct. The use of the CRIMMs does indeed introduce some extra latency to the RDRAM mix. Again, it isn't a huge amount, but the difference is definitely there. If you look at those MPX/DDR numbers, it makes you wonder why anyone ever complained about RDRAM's latency to begin with. Heh.

Linpack and Cinebench >>