Many of the best-performing accelerators we tested had only beta Windows NT drivers during our test cycle. In fact, the only shipping NT drivers at the time of testing seemed to be those that ship with NT itself. Aside from the drivers with results shown in the accompanying bar chart (all shipped on the Windows NT CD-ROM), we received beta drivers for the SuperMac Spectrum/24 VL for Windows, Diamond Viper VLB, Hercules Graphite VL Pro HG720, and Matrox MGA Impression/3/V.
At this early stage, the low-cost Hercules Dynamite VL pro D602 walks away from the field (but it is still slower than it is in Windows 3.1). The NT tests ran on a 486/66 DEC MTE.
For Windows NT testing, we used software based on the same code as the Windows tests with a few co
nditional compilation switches. Testing graphics performance under NT can be tricky, thanks to NT's automatic GDI (Graphical Device Interface) instruction batching and dictatorial grasp of memory.
The instruction batching seeks to minimize the overhead of making many small requests to the graphics driver by grouping many small requests into one big request. Carelessly constructed tests can end up measuring the time it takes to place an instruction into this cache, rather than the time it takes to paint the screen.
The memory problem stems from NT's refusal to lock more than 128 KB of a process into memory. This limitation causes problems when testing adapters with 4 MB of VRAM. An unintentional swap can disrupt results.
Chart: Accelerator - Response time (seconds; 256 colors)
Windows NT Drivers: Better Late Than Never?
