Tim Sweeny forecasts end of GPU

[xadhoom]

Here are the slides:

http://graphics.cs.williams.edu/archive ... PG2009.pdf

[milliams]

An interesting presentation. I note that he mentions volumetric rendering using things like marching cubes explicitly as a technology of the future.

[Borundin]

Yes, very interesting especially from page 38 and onward.
We had a somewhat related discussion a few years back here on the forum. Of course that was before Larrabee and other recent hardware was released.

[jjp]

Sweeny has a weak spot for software rendering. This isn't the first time he forecasts the end of the GPU. On this subject I don't value his opinion very much.

EDIT: although in the sense that we may one day give a different name to what we now call 'GPU' he is probably right.

[_tommo_]

really interesting, i came to exactly the same conclusions after trying to implement these principles (no DX) in my CUDA render.

In particular i quote his idea that GPGPUs will beat APIs because it's easier, not faster... and that the current huge GPU memory bandwidth is really small compared to the needs of REALLY generic data-parallel algorithms.
In fact, implementing the rasterizer was a 0-tricks task, completely different from implementing something in DX...
the trickery involved beating the memory bottleneck for rasterizer writes.

Also, it's true that a successful "general compute" model HAS to have shared memory beween sequential/streamed (CPU/GPU) processors... or it is just too slow.
Strangely, because of this the last laptop and mobile architectures would be the best ones for GPGPU (CPU and GPU on the same die), if they had enough processing power.

Anyway "GPUs" will never die... in fact absorbing all the compute properties of the CPU they will become a central part of any system, even if under another name.