Performance Testing with FX!32

The Windows NT Magazine Lab ran performance tests on aprototype FX!32 system from Digital. The system had a 466MHz Alpha 21164A(version EV5) and 128MB of RAM (see "New Alpha PC Strategy," page140), running Windows NT 4.0 Workstation (Release Candidate 2.3, Build 1362) anda beta release of FX!32.

We expected FX!32 to perform at 20% to 30% of native and this new superhigh-speed chip to outrun everything we'd tested on Alpha-native programs (see "NewAlpha CPU Raises the Bar," page 144). In some cases (such as on Photoshop),the translated code's speed blew us away. In other cases (such as LightWave 3D),the performance disappointed us. Unfortunately, getting an accurate read of howfast Photoshop really ran relative to native performance is impossible becausean Alpha version does not exist. For a baseline comparison, we ran ourbenchmarks on a high-end graphics PC from NeTpower (a Calisto with a 200MHzPentium Pro, 128MB of RAM, and a Matrox Millennium graphics card with 2MB ofWRAM).

Before I get too deep into the numbers, let me explain a few more detailsof why FX!32 can run so fast, or why it might not. On the plus side, FX!32 isfully multithreaded, is OLE2 aware, and has full support for the Win32environment (FX!32 can use native APIs, drivers, and other code), and theruntime interpreter is small enough to fit into available CPU cache forincreased performance. We optimized the Intel code as much as possible beforecounting the test values: We ran the applications several times on the functionswe were testing and turned the optimization parameters all the way up.

Granted, we ran a beta release of FX!32 and not the fully tuned versionthat came with NT 4.0. Running a beta emulator on a beta operating system on aprototype computer is bound to cause a few problems. I expect that a fullrelease version of NT 4.0, FX!32, and a production-level computer will performbetter.

Table A lists LightWave 3D 4.0's render times for several types of images(ray traced, texture mapped, etc.) in the native Alpha versions, the translatedIntel version, and the native Intel version on the NeTpower. As you can see, wewere at about 10% to 20% of native performance on average. Table B shows scriptruntimes for our tests on graphics workstations: a simple 2D morph in ElasticReality and a test script in Photoshop, which manipulates a 28MB image file.