The State of Linux NVIDIA Overclocking

Having only played around with NVClock a little prior to the writing of this article, we decided to first try NVClock 0.7 on one of our old Riva TNT2's. From the years of owning this graphics card, we've given it more than a couple of beatings and the card had served its purpose to us. A simple passive aluminum heatsink rests atop the TNT2, but we had installed an Antec VCool in the expansion slots immediately below the AGP slot to assist in the cooling process. The memory found on the Riva TNT2 was Winbond 943WC W986416CH-7. This memory is rated for 143MHz/CL3 with a clock cycle time of 7 ns.

With stock speeds of 125/143 (VPU/MEM) the TNT2 is undoubtedly sub-par by today’s standards. But the TNT2 is generally known to be a good overclocker, and we have achieved overclocking success with this and other TNT2 cards back in the days of running Quake 3 and Wolfenstein. When loading up NVClock, the frequencies read 125.007/143.180. We proceeded with running the benchmarks at stock speeds, so we had some quantitative results to compare. Our stock results for Unreal Tournament 2003 were a whopping 11.6 average FPS for dm-asbestos and 8.1 for ctf-citadel. The Unreal Tournament 2004 frame-rate were worse, as expected, with ons-torlan having an average of 3.4 average FPS and dm-rankin at 6.2. For the fun process to begin, we first decided to see how far we would be able to push the Winbond memory from its stock speed of 143MHz as stated Winbond Electronics W986416CH datasheet. For the memory overclocking in NVClock, we proceeded to raise the memory clock in about 5MHz increments. When we hit 125/185 is when the screen turned to yellow and purple lines, so we had backed down to 180MHz where there were no visible artifacts. We benchmarked Unreal Tournament 2003’s dm-asbestos to check for artifacts and to see the change in average FPS. About a 2.5FPS increase had occurred with this card with a score of 14.1. As there were no artifacts at these speeds, we proceeded with bumping up the VPU frequency. With the same technique as we had used for increasing the memory speed, we managed to increase our core frequency to 179MHz, going anywhere past this point would result in artifacts. Now finding the threshold for this card, we proceeded to fine-tune these frequencies. However, when attempting to adjust the memory and core frequencies further, artifacts began to appear. After several more rounds of artifact testing in gaming at 179/181 the graphics became littered with artifacts and instability issues so we ended up backing down to 166/172 after testing it at several other higher speeds.