[cairo] Re: [Dri-devel] Gains from doing 2D drawing using OpenGL/mesa

[Alan Cox]

On Mer, 2004-04-21 at 21:32, Jon Smirl wrote:
> I'm not sure we'll ever see that. Memory bandwidth is the problem, not CPU/GPU
> processing power. You want to keep the setup with the GPU to minimize the memory
> load on the main processor.

All the low end solutions are UMA. Seperate video memory is fine for the
less cost sensitive end of the market. UMA makes host CPU render-assist
work (it doesnt work over PCI like busses for latency reasons)

You also do need the assist (especially on x86). If you benchmark
software mesa you can see that.

> The hyper-transport link would let both processors get at the
> video memory at full speed but also keep the GPU traffic from interfering with
> the main CPU. If the GPU runs out of memory it could transparently spill over
> into system memory.

On Athlon64 there is no "main memory", its all owned by hypertransport
devices (currently the CPUs). A HT GPU could probably also be a memory
owner I guess.