So what is next for the GPU?

[rwolf]

I for one sure hope physics never ends up being done on the GPU, because it is quite bogged down as it is drawing graphics (and it's unlikely things will lighten up in the future).

With almost all customers unable to buy higher-end graphics parts, it wouldn't help anyone if part of the already mediocre graphics performance was chewed up running physics simulations. Instead, bring on the dedicated silicon I say. At more reasonable prices than have been suggested so far, I might add.

I can see the performance issue you are raising and it is valid however just image with a PPU you have to transfer each frame over the bus. It would be better to put the ppu chip on the graphics card sharing memory with the GPU.

 

[DemoCoder]

Scene graph acceleration. Nuff Said.

 

[JasonCross]

Any ideas about the future direction of the GPU? What is the next big thing?

I suspect that we will see things like lighting, shadowing, 3d clipping, and physics move to the GPU. Unified pipelines will be more signifiant more for the application code they can or could run than for the advancement of pixel and vertex shaders.

The next "big thing" is DirectX 10 compatibility.

Unlike DX9, you can't implement only part of the DX10 spec. It's all or nothing - there are no "cap bits" in DX10. This is great for developers...if they want to use shaders a certain way, or a particular blend mode, or whatever the case may be, they have a fixed platform to write to. No more of this "different shader path" nonsense.

No more different results for the same operations, too. DX10 defines strict compliance margins of error for shader math and pixel output. If you submit API code, the hardware has to give you the right result (within a very small margin) or you don't have DX10 hardware.

DX10 also changes up the pipeline a bit, with things like a scheduler and geometry shaders unit (in addition to vertex shaders). Also stuff like state buffers and so on. It's a pretty big change.

So making DX10 compliant hardware is going to be job #1 for ATI and NVIDIA for the next year or two.

Virtualized memory access is coming to DX10 as well, though possibly not when it launches with Windows Vista, right now it's "shortly after." Don't know if the hardware guys are going to implement it sooner - certainly there are some good things you can do driver-level with full memory virtualization.

I can see the performance issue you are raising and it is valid however just image with a PPU you have to transfer each frame over the bus. It would be better to put the ppu chip on the graphics card sharing memory with the GPU.

You don't really have to transfer entire frame info over the bus. There's a relatively small subset of positional data that needs to be transferred to do physics really well - the data that a PPU needs to manage is huge, but the frame-to-frame data that needs to come into and go out of it is relatively small. That's one reason why the first AGEIA cards will be PCI and not x1 or x4 PCIe...it's just not strictly necessary. Of course, more bandwidth will help, but it's not the primary reason to move it to the GPU.

 

[rwolf]

Scene graph acceleration. Nuff Said.

Now were talking.

http://www.gamedev.net/reference/programming/features/scenegraph/

http://www.openscenegraph.org/

 

[rwolf]

JasonCross, hopefully DirectX 10 will reduce the cost and improve the quality of games. Standardization is great from that point of view.