Single page Print

SPECviewperf
Now this should be a little more up the P750's alley. SPEC's viewperf simulates graphics loads taken from real-life, workstation-class OpenGL applications.

Anyone looking at using the P750 in a low-end CAD workstation should be a little disturbed by the card's relatively slow performance in SPECviewperf 7.1. It's embarrassing to see the P750 and even the high-end Parhelia getting whupped by ATI and NVIDIA's bargain basement graphics cards.

However, these SPEC results don't follow the performance pattern we've established throughout our 3D testing. Curious, I asked Matrox about the low scores, and they sent me the results of their own internal testing with an nForce2 system and Athlon XP 2400+. Matrox's scores are below:

  P750 Parhelia
3dsmax-026.6658.938
drv-0916.9717.21
dx-0837.5842.34
light-060.0009.024
proe-023.3964.465
ugs-035.1636.023

Matrox's internal scores for the P750 and Parhelia are closer to what I would have expected from the cards, and they are much higher than the scores our test card was able to achieve. After several fresh drive images, driver installations, and clean installations of Windows XP, I was unable to replicate Matrox's scores on our nForce2 test system or on a couple of Pentium 4-based testbeds.

I've been working with Matrox on this problem for a couple of weeks now. While my test results in other applications suggest that Matrox isn't cooking the books intentionally, I can't see end users having better luck getting their systems to match Matrox's internal scores than I did. Short-Media's review of the P750 also shows poor SPECviewperf performance, which only reinforces my confidence in the repeatability of my scores. I'm still working with Matrox on the problem, and I will update this review if Matrox is able to fix whatever's wrong.

You may also be curious about why the Millennium P750 fails the light-06 test while Parhelia produces a score. Since the two cards have the same feature support, they should both be able to complete the test. According to Matrox, the P750's failure has to do with the fact that it only has 64MB of memory; SPECviewperf apparently tries to write to areas of memory that don't exist on the P750. However, our 64MB Radeon 9000 Pro has no problems with the light-06 test, so I suspect the P750's drivers are at least partially to blame.

Update 24-11-2003: Matrox has released a driver update for its Parhelia, P750, and P650 graphics cards, fixing a nagging bug that impeded performance in SPEC's viewperf workstation graphics benchmark. The P750's performance with the latest 1.05.01.008 drivers is much improved and the card generates the following scores on our Athlon XP 2600+ test system:

 P750 (1.05.01.008)Parhelia (1.05.01.008)
3dsmax-026.0988.789
drv-0915.2417.10
dx-0835.7842.19
light-068.7048.911
proe-023.3564.356
ugs-034.7236.042

As you can see, the new drivers also fix a bug that prevented the light-06 test from completing on the P750.

AGP texture download performance
Because the P750 is suited for video editing, we also threw in Serious Magic's AGP texture download benchmark. Serious Magic develops video editing applications, and decent AGP texture transfer performance is key to certain video manipulation tasks. You can read more about this issue here.

Since we first looked at texture download speeds, driver updates from ATI and NVIDIA have dramatically improved the performance of the Radeon and GeForce cards, but Matrox's cards are still slow. The P750's poor texture download performance will be an issue for any application that needs to bring the card's output back into main memory. Even simple 3D character generation or screen captures will be limited to very low frame rates and resolutions. Since ATI and NVIDIA were able to improve their cards' texture download performance dramatically with driver updates, we'd like to see Matrox address the problem with a future driver release.