LOSTCIRCUITS

SPEC.org vierwperf 8.1

We briefly brushed upon the issues with viewperf in the beginning of the article, that is, the outdated software versions that are single-threaded only and so on. Nonetheless, viewperf is an interesting benchmark geared towards high end professional graphics cards even though it is oftentimes abused to run with consumer cards - for whatever that's worth.

The detailed viewset descriptions are available here, for easier digest of the results we show screenshots of each application (courtesy ot SPEC.org) .

For testing, we used a different approach, in short, we were interested in whether there were any differences in the SLI interfacing, that is, whether there are differences between the nForce4 Pro with its 2 x16 lanes SLI configuration and the standard SLI with 2 x 8 lanes. We also tried to look at the impact of different CPUs, that is single vs. dual core Opterons and different clock speeds for both CPU and, conversely, for the GPU / frame buffer of the graphics cards to pinpoint the bottlenecks.

3dsmax-03

Bugs: with any but the 83.xx drivers we experienced a rather strange bug: Unless there was a separate USB device plugged in (we were using a logitech wireless mouse), we found a bimodal distribution of average frame rates with one set around 25 fps (average) and the other one aroud 45-50 fps. The solution to the problem was that as long as we manually kept wiggling the mouse, the system remained in high-speed mode, as soon as the mouse was left idle, the system defaulted into "slow mode". At this point, it is stil not clear what actually caused this issue but it was 100% reproducible even throughout several format c: and reinstallation of the operating system, drivers and applications.

Frames per second, higher is better. The maroon columns show the Foxconn board, the blue columns represent the DFI board with the multipliers given for each CPU at a bus clock of 200 MHz. The -2898 moniker indicates running at a 14 x multiplier and 207 MHz bus speed for 2989 MHz clock speed. The "OC" suffix indicates overclocking of the Quadro FX4500 to 500/1300 MHz [core/memory]. Pink and rose columns represent running in a WoW under Windows XP-64 at 2898 MHz CPU clock and standard / overclocked graphics adapters.

There are a couple of things that stick out:

• dual cores perform better than single cores
• the nForce Pro is clock for clock slightly faster than the nForce4 SLI (with the exception of the 13 x multiplier setting)
• Overclocking the graphics cards did not improve performance
• Under WindowsXP-64, the system runs noticeably faster than in WinXP-32

catia-01

3dsmax was the only application sensitive to the "mouse-bug", no other application was affected

Frames per second, higher is better. The maroon columns show the Foxconn board, the blue columns represent the DFI board with the multipliers given for each CPU at a bus clock of 200 MHz. The -2898 moniker indicates running at a 14 x multiplier and 207 MHz bus speed for 2989 MHz clock speed. The "OC" suffix indicates overclocking of the Quadro FX4500 to 500/1300 MHz [core/memory]. Pink and rose columns represent running in a WoW under Windows XP-64 at 2898 MHz CPU clock and standard / overclocked graphics adapters.

Summary:

• The nForce4 SLI is clock for clock slightly faster than the nForce Pro
• Overclocking the graphics cards did not improve performance
• Under WindowsXP-64, the system runs noticeably faster than in WinXP-32

Next Page:    => Ensight-01, Light-07 =>