Nvidia’s new 20-series video cards, due to ship next week, will introduce a new overclocking system: Scanner.
Overclocking is typically a trial-and-error process: increase the clock speed, run some intensive workloads to make sure everything works OK, maybe fiddle with the GPU voltage to eke out a bit more stability. Push the clock speed too high, and the system stands a good chance of locking up and crashing.
Scanner will make this process a lot more automatic and trustworthy. It will run incrementally and increase the clock speed and voltage of the GPU to build a profile of a specific card’s capabilities, testing the GPU’s computing ability at each speed to make sure that it’s operating properly. Generally there will be arithmetic errors, rather than outright crashes, when the GPU is being operated only very slightly faster than it can support. So by making only small adjustments, Scanner can probe the limits of what the chip can handle without the crashes and reboots, and without human intervention. The whole process takes about 20 minutes.
The overclocking performed by Scanner only applies to the GPU; for memory overclocking you’ll need to stick with traditional methods. If you prefer to use those traditional methods for the GPU itself, that’s still going to be possible, too.
Scanner isn’t an application; rather it’s an API. Scanner support will be integrated into the various overclocking apps built and shipped by GPU vendors. At launch, it will support only the new 20-series parts, but support for older chips is expected to ship later.
This is a smart approach to overclocking, so much so that it’s a bit surprising that it’s taken so long to arrive. Nvidia knows how best to induce hot-spots and thermal instability in its chips, so in principle at least, Scanner’s determination that a given speed is stable should be a lot more accurate than the kind of “3DMark runs OK so it’s probably good” testing that end-users normally use. Because it’s automated, Scanner can make smaller incremental changes than a human would, without being overwhelmed by the tedium of the trial and error testing. One wonders if all overclocking software shouldn’t work the same way.