Speeding Up The Linux Kernel With Your GPU

[matthewbpt]

Well it's simple:
1. nVidia hands out no documentation, so you need the blob;
2. Open source GPL parts refer to the blob;
3. Linux depends upon proprietary.

Yes, except no parts of the kernel need, or are likely to need, CUDA to run. This just gives one the option of running kernel code on GPU, there's no chance that the kernel will ever move to requiring a GPU for kernel level operations.

[V!NCENT]

Seeing as the linux kernel is GPL'ed I don't see how this could be possible ... no conspiracy theories please ...

Well it's simple:
1. nVidia hands out no documentation, so you need the blob;
2. Open source GPL parts refer to the blob;
3. Linux depends upon proprietary.

[matthewbpt]

Just to put this into perspective, this project is all about nvidia trying to turn the linux kernel into their own personal proprietary blob. Drop dead nvidia!

Seeing as the linux kernel is GPL'ed I don't see how this could be possible ... no conspiracy theories please ...

[V!NCENT]

Just to put this into perspective, this project is all about nvidia trying to turn the linux kernel into their own personal proprietary blob. Drop dead nvidia!

No. nVidia needs Linux to manage their hardware. Don't forget that they are working on a CPU that can never match AMD or Intel. Linus would never accept it and nVidia knows this.

Linux is key to their hardware adoption in this regard and therefore they can't and won't do that.

[droidhacker]

Just to put this into perspective, this project is all about nvidia trying to turn the linux kernel into their own personal proprietary blob. Drop dead nvidia!

[V!NCENT]

You are arguing but saying the same thing. SSE and the likes brought vector specific registers to the x86 much like Altivec did for the PPC. In fact AVX is an effort to further improve on those capabilities.

Wasn't the thing we were arguing about that the CPU offloads these calculations to the GPU with a standardised instruction set instead of replacing them with the CPU instruction sets?

I thought you were saying that Larrebee ofloaded them. I meant to say that it does them.

[deanjo]

Lol, no... Larrabee was some sort of CPU design that allowed you to do vector calculations in the CPU. These vector registers could aid in 3D rendering, but that's about it, basically.

I still have that Intel paper somewhere in my Gmail account is case you don't believe me...

You are arguing but saying the same thing. SSE and the likes brought vector specific registers to the x86 much like Altivec did for the PPC. In fact AVX is an effort to further improve on those capabilities.

[V!NCENT]

Yup it could in theory but it would be much slower and inefficient. That is basically what intels Larrabee was trying to do.

Lol, no... Larrabee was some sort of CPU design that allowed you to do vector calculations in the CPU. These vector registers could aid in 3D rendering, but that's about it, basically.

I still have that Intel paper somewhere in my Gmail account is case you don't believe me...

[deanjo]

Silly question:
Can't GPUs be made to assist the CPU in software rendering (3D, video) via a standard instruction set extension (say 'SSE42' or 'x88')? Wouldn't that allow us to get rid of all the 'graphics driver' mess and have such things HW accelerated independently of the specific hardware?

Yup it could in theory but it would be much slower and inefficient. That is basically what intels Larrabee was trying to do.

[not.sure]

Silly question:
Can't GPUs be made to assist the CPU in software rendering (3D, video) via a standard instruction set extension (say 'SSE42' or 'x88')? Wouldn't that allow us to get rid of all the 'graphics driver' mess and have such things HW accelerated independently of the specific hardware?