Friday night topic: When CPUs and GPUs converge

— 5:15 PM on April 19, 2002

It's inevitable. CPUs and graphics chips are on a collision course. As graphics chips become ever-more programmable, the work they do—processing vertex and pixel shader instructions—looks more and more like the work done by a Pentium 4 or Athlon. APIs like DirectX and OpenGL are being rearchitected to be more general—to focus less on specific, traditional graphics functions and more on shader programs that manipulate graphics data.

Meanwhile, CPUs gain ever-more memory bandwidth and specialized SIMD instruction sets, which make them look more like GPUs. The Pentium 4 has an especially impressive front-side bus and the power to move gobs of data across it. AMD's upcoming Hammer eliminates the front-side bus altogether, tunneling AGP data to an AGP controller chip via high-speed HyperTransport links. Instead of a north bridge chip, Hammer includes an integrated memory controller for better memory performance.

Eventually, Intel will incorporate pixel and vertex shader instructions into a future iteration of SSE, or AMD will do something similar with a SIMD instruction set. Probably both. (Already, DX8's optimized software vertex shader routines do a decent job of keeping up with the action.) When that happens, it'll be armageddon for graphics chip companies. The GPU will risk becoming just another unit on the main processor, the x87 of the new century. Fabless wonders like ATI and NVIDIA will have to contend with the mighty manufacturing prowess of CPU makers, and it won't be pretty.

Or maybe the trend toward custom ASICs handling specialized tasks will continue as it has been going for years now. I dunno.


Tip: You can use the A/Z keys to walk threads.
View options

This discussion is now closed.