View Single Post
  #2 (permalink)  
Old 31st March, 2004, 04:29 PM
Aedan Aedan is offline
Chief Systems Administrator
Join Date: September 2001
Location: Europe
Posts: 13,075

The GPU and other components are driven from clock signals derived from a crystal. As the fundimental frequency of this crystal is too low, it is multiplied upwards by a PLL. As the PLL can be programmed to provide a particular frequency, all the software tools do is change the multiplication used by the PLL.
Hence, to overclock a GPU or memory, you just have to reprogram the PLL's internals.

Now to the other part of the problem! It's not the moving of the electrons - they'll move when you apply a potential difference across them. The problem is making out the difference between a 1 and a 0 when the data gets to the end of a transmission line. Increasing the voltage to the I/O buffers helps to increase the difference between a 0 and a 1, and hence increases the noise margin.

Heat causes a problem too, as the hotter a component gets, the larger the thermal noise that is generated. Once this noise gets to the point it could be misintepreted as a 0 or as a 1, then the game is over. This happens easily, as at RF frequencies, the I/O buffers will be seeing a bunch of cascading edges, rather than discrete voltage levels.

Once you look into high speed logic design, you begin to realise that it's amazing anything works at all!
Any views, thoughts and opinions are entirely my own. They don't necessarily represent those of my employer (BlackBerry).
Reply With Quote