ATI, AMD, and free drivers
Posted Aug 3, 2006 22:59 UTC (Thu) by
bluefoxicy (guest, #25366)
In reply to:
ATI, AMD, and free drivers by NRArnot
Parent article:
ATI, AMD, and free drivers
Graphics go on the north bridge, the high-bandwidth bus between the CPU and the memory/AGP/PCI-E16x/etc hardware. They do not belong on the CPU, ever.
Having CPU + GPU on the same chip means a strong lack of flexibility. The die has to be redesigned to reposition the transistors to fit either A) A new CPU core; or B) A new GPU core; this is -very- -expensive-. The north bridge usually connects the GPU to the PCI-Express or AGP bus; in effect, the north bridge stays the same, the GPU is effectively run to a controller chip for a PCI-E card which is just wired straight into the north bridge. This gives cheap and easy flexibility because you swap two chips on the board if you want to release a new board with a better GPU.
Besides that, both GPU and CPU activities cause heat. You don't want to concentrate that in the same chip. Imagine a Barton running twice as hot; instead of a 2.0GHz Barton you have a 1.2GHz Barton to control the heat. Worse, GPU and CPU are independent, so you would have temperature fluctuations based on two separate models.
As for plugging into another CPU socket, CPU sockets are all proprietary. You don't want to search for an ATi Socket A vs ATi Socket 754 vs ATi Socket 939; this puts more manufacture stress on the company to satisfy the same market. They either have to manufacture all kinds of chips; or reduce their target market. Both are bad for business and will increase costs and thus consumer prices. Then you also have the issue of changing connectors, i.e. D-SUB vs DVI vs whatever they come up with nex; how do you change those? Buy a new mobo?
Single-chip CPU/GPU is too expensive and inflexible to work. It's not that it brings substantial gains but has a few hurdles; it's that it brings substantial problems and has negligible gains. Huge bandwidth between the GPU and the CPU is needed to load textures; besides that we're sending just a ton of simple commands. Video memory and video card DMA (which is kind of what AGP did) handles the textures and models, since they can be loaded in and sit waiting to be used; CPU-GPU integration will make the commands go more realtime by maybe 1/10000 FPS, not useful.
(
Log in to post comments)