First post, by vvbee
- Rank
- Oldbie
On a K6-2 300,* the Voodoo 3 2000's performance is bottlenecked, as is common knowledge. At resolutions up to 800x600, the game Need for Speed 5 (below) at medium settings is capping at an average of about 35 FPS running a 120-second replay of a race on the game's first track regardless of whether the Voodoo 3's core and memory are clocked at the default 143 MHz or underclocked to 100 MHz.
In that game, at 640x480, a paired t-test suggests that performance at 143 MHz does differ from that at 100 MHz (p = .022); but the figure below (left) shows that this framerate, sampled once per second over the duration of the replay (n = 120), isn't meaningfully distinguished between the two operating frequencies, differing on average by 1 FPS (triangles = FPS samples at 100 MHz, circles = FPS samples at 143 MHz). At 800x600 (middle), there's a slightly more tangible benefit for 143 MHz (highest 10% FPS average = 46) over 100 MHz (highest 10% FPS average = 44), but given that the latter is a 30% underclock, a 5% drop in performance isn't notable. At 1024x768 (right), there's a more clear-cut benefit, mostly across the FPS spectrum, for running the card at 143 MHz rather than 100 MHz. The framerates given here are truncated averages of two runs each.
A similar trend is found running Quake 3 (DEMO001; default settings; left graph below) and 3DMark 99 MAX (race test; right graph below). Up to and including 800x600, there's on average no framerate benefit to running the Voodoo 3 at 143 MHz: even at 92 MHz, the lowest underclock my card would do without memory issues,** no real differentiation in speed is seen. The framerates given here are truncated averages of two runs each.
The benefit of underclocking is a reduction in core - and possibly RAM - temperature. Underclocking the Voodoo 3 from 143 to 100 MHz, I found a 3 °C reduction in core idle temperature, from 43 to 40 °C (against 24 °C in-case ambient),*** and the same result was had with a repeat some hours later. At the end of a 10-minute loop of DEMO001 and DEMO002 in Quake 3, the temperature was 46 °C at 143 MHz and 42 °C at 100 MHz. The temperatures were measured running at 143 MHz first, then at 100 MHz, which by design may have given a slight edge to the former. It's likely that the GPU core was running at notably higher temperatures than those measured at the surface of its heatsink, as done here, which means that the relative differences in temperature would likely be found larger if measured directly from the core. Quake 3 had a resolution of 800x600, here, i.e. running bottlenecked.
In short, when using resolutions of 800x600 or less, and certainly 640x480, on this system with the programs tested here, there's in general no performance penalty for underclocking the Voodoo 3 from 143 MHz to 100 or even 92 MHz, and doing so will reduce the card's operating temperature and possibly increase its lifetime.
* Test system specs: AMD K6-2 300 CPU (100 MHz FSB), PC-Chips M577 motherboard (Csongrádi/Steunebrink 990309J2 BIOS), 64 MB PC-100 RAM, Voodoo 3 2000 AGP GPU, IBM Deskstar 25GP harddrive, Windows 98 SE.
** The card's RAM chips are labeled SEC KM416S1020CT-G7. I can't decipher from their datasheet what the minimum frequency is, but below 92 MHz, I get garbage on screen.
*** As measured with a digital contact thermometer from roughly the middle of the top of the core's heatsink, after the PC had spent 10 minutes idling on the Windows desktop. Throughout, the heatsink received ambient-temperature airflow from a 45 CFM case fan ~25 cm away.