Want More Info About Vector Processors

[PixelOz]

Does anybody know if these new vector processors in modern GPUs are similar or comparable to those vector processors that some supercomputers had (like those in some Cray supercomputers)?

If so can you shed more info about their differences and about their similarities?

 

[Simon F]

I'm not familiar with Crays, but in my Uni days, I was asked if I'd be interested in rewriting FORTRAN code** for a, IIRC, a Cyber 200 machine which (again, IIRC) could do > 200MFLOPS! (This was in the early 80s).

To make use of the vector hardware, it assumed you could rewrite code so that you had, for example, sets of long arrays of floats and the program to do operations in a form equivalent to, say,...

for(i=0; i < ARRAY_LENGTH; i++)
{
    c[i] = a[i] + b[i];
}

The compiler then had 'libraries' with functions that did operations like the above using, say, a single instruction.

This is not like the shader models of graphics chips which assume you can do ops on fixed 4-element structures.

**It was a holiday programming job. In the end, I was very glad I didn't get it as I spoke to the guy who did and he said it was not much fun

 

[Dade]

Does anybody know if these new vector processors in modern GPUs are similar or comparable to those vector processors that some supercomputers had (like those in some Cray supercomputers)?

If so can you shed more info about their differences and about their similarities?

I have used, a lot of time ago, a CM: http://en.wikipedia.org/wiki/Connection_machine

It was a wonderful piece of hardware and it was programmed in C*: http://en.wikipedia.org/wiki/C*

Now I have only a vague memory of how it was working but it was a pure SIMD machine compared to modern GPUs: it was capable to perform operations in parallel on vectorial data types however there was no direct support for the execution of conditional statements, etc.

Instead, modern GPUs are SIMT (Single Instruction Mutiple Threads) and are capable to execute contitional statements, loops, etc. (by masking results). They are a lot easier to program.