News for hackers. Stuff that matters.
Re:NVIDIA Announces Tesla Personal Supercomputer by wmf on Wednesday November 19, 2008 @05:17PM (#370009)
If all threads in a block have the same number if iterations, you're fine; different blocks can take different number of iterations. As long as the number of blocks is much higher than the number of SMs, the machine will dynamically schedule the blocks to available SMs.

At this point the programmer's head has already exploded. MIMD systems (like OpenMP on Larrabee) don't have any of this BS.

Re:NVIDIA Announces Tesla Personal Supercomputer by miloshh on Wednesday November 19, 2008 @07:33PM (#370143)
No, Larrabee will have exactly the same issues. Except threads will be called fibers, and blocks will be called threads. Read the Larrabee paper from Siggraph 08.

The way to get maximum performance out of a given chip area is to use SIMD, and that's here to stay, with all the associated issues.

Re:NVIDIA Announces Tesla Personal Supercomputer by wmf on Wednesday November 19, 2008 @08:37PM (#370201)
Except threads will be called fibers, and blocks will be called threads.

That's how their rasterizer works, but I'm talking about using it as a regular x86.

The way to get maximum performance out of a given chip area is to use SIMD, and that's here to stay

There's a big difference between MIMD+narrow SIMD and super wide SIMD.

Re:NVIDIA Announces Tesla Personal Supercomputer by miloshh on Wednesday November 19, 2008 @10:04PM (#370257)
The Larrabee SIMD width will be 16 and NVIDIA's is currently 32, so that's almost the same.

Not just the rasterizer, but any application that wants to take full advantage of Larabee will have to use the SIMD vector units to the max.

Larrabee might turn out to have great performance (which I hope), but if it does, the reason will not be black magic or breaking laws of physics. The reason will be SIMD.

Re:NVIDIA Announces Tesla Personal Supercomputer by wmf on Friday November 21, 2008 @04:03PM (#372330)
For some reason it's easier for me to wrap my head around one thread driving a 16-wide SIMD unit than 32 threads that execute in lockstep. I know it ends up being equivalent but it feels different.

Also, on Larrabee you can execute a different kernel on each core, while on GPUs you can't.