marius popa wrote:

>I saw it on this page http://www.gpgpu.org/s2004/ and it gives me some
>vague ideas :)
>Let's say some tasks could be made on GPU : blob filters ,searching
>trough images :
>
>select from * from pictures_blobs where image like "darth_vader.jpg"
>
>
>
>

The slide show suggests that more cycles are better than few cycles.
The question is whether you get more bang from more cycles from general
purpose ("smart") processors or special purpose graphics or signal
("dumb") processors. Smart processors are easier to schedule, more
versatile, easier to run parallel, etc. It also easier to allocate
memory when all processors equal access.

A processor can be made faster by increasing the clock rate, pipeling
complex instructions, and superscalar execution. Modern general purpose
processors do all three plus supporting cache coherency for
multi-processor execution. The special purpose processors generally
have a high cycle rate of simple instructions.

This is essentially the risc vs. cisc argument. The risc guys made the
best argument but are being tossed on the scraphead of technical history
by pure, unadulterated cisc technology. I've always been an agnostic on
the risc vs. cisc battle, but I have to admit that 3+Ghz superscalar
cisc processors blow me away.

Dell is now advertising a 512MB, 2.8GHz 64 bit Xeon processor, SMP
ready, with dual 80GB SATA drives and Raid-1 for $599. Is a $400 video
board going to make it faster?


--

Jim Starkey
Netfrastructure, Inc.
978 526-1376