Larrabee architecture

[jjp]

So.. we're back at software rendering or what? I'm still sceptical about the performance but the architecture looks very interesting: http://softwarecommunity.intel.com/User ... nycore.pdf

[syd]

as current graphic cards are single core, I can't see anything else than huge performance increase
Just imagine each single element of the scene being rendered in its own thread...

[jjp]

as current graphic cards are single core, I can't see anything else than huge performance increase
Just imagine each single element of the scene being rendered in its own thread...

Uhm, current graphics cards are massively "multi core" They run hundreds or thousands of threads at a time. I doubt the first generation of Larrabee will offer more than mainstream cards from ATI/NVidia at that time. But it probably will be a lot better at GPGPU tasks.

[syd]

thanks for the info

[triton]

http://www.anandtech.com/cpuchipsets/in ... spx?i=3367

I found that article pretty interesting.

[Kojack]

Assuming they can break the Intel gfx chip tradition (of being fairly useless), it's damn interesting even if it wasn't quite as fast (but close). As the doc mentions, some of the fun stuff you can do easily include shaders directly reading from render targets for user defined blending modes and single pass tone mapping, order independent transparency by storing per pixel spatial data structures, and the fact that it can run c++ code using intel's auto vectorisation compiler.

The possibilities are pretty exciting (assuming that it doesn't suck).

The fact that Tom Forsyth is on the team and Michael Abrash helped design the instruction set is enough for me to be optimistic.

[Klaim]

Doesn't x86 allowed on gpu assure that that GPU will not be as fast as have we have currently?

(it's a real question, i have no idea)

[xavier]

Assuming they can break the Intel gfx chip tradition (of being fairly useless), it's damn interesting even if it wasn't quite as fast (but close). As the doc mentions, some of the fun stuff you can do easily include shaders directly reading from render targets for user defined blending modes and single pass tone mapping, order independent transparency by storing per pixel spatial data structures, and the fact that it can run c++ code using intel's auto vectorisation compiler.

The possibilities are pretty exciting (assuming that it doesn't suck).

The fact that Tom Forsyth is on the team and Michael Abrash helped design the instruction set is enough for me to be optimistic.

It's actually the same architecture (in theory) that consoles use -- shared memory instead of dedicated VRAM, which allows for more interesting things such as writing GPU commands and submitting rendering resources directly from the L2 cache.

The problem is that they are trying to make it work with the GPU-based rendering APIs, and those API's are not really designed to work this way -- the Direct3D used on the 360 is not the same Direct3D used on PCs.

[jjp]

Doesn't x86 allowed on gpu assure that that GPU will not be as fast as have we have currently?

The fact that it does have the x86 ISA will cost some transistors but it doesn't impact performance as such.