Jared Putnam wrote:
The weaknesses of GPUs are:
There is no support for integers or fixed point.
Actually, this isn't a weakness. IIUC, GPUs use unsigned integers only
for control functions. Is there really a need for signed integers?
And, AFAIK, fixed point only exists in software.
There is only fixed-function graph reduction. Output must be a
two-dimensional array. There is no hardware support for doubles,
although there are SIGGRAPH presentations this year about emulating
them.
So, a SIMD processor that executed a program would have advantages.
If those can be beaten in a product costing less than the Cell, there
will be markets.
My somewhat excessive design spec was for registers which held 4 32bit
floats and 4 hardware floating point multipliers that could be used with
FMACs to do a [4,4] * [4,4] matrix multiply in 4 clocks and a four input
floating point adder to do [4,4] x [4,4] matrix multiply in 16 clocks.
Scientific computing wants cheaper supercomputers,
financial and statistical computing wants faster APLs,[1] game
programmers[2] want to run "essentially functional" programs which
transform "a small input data set to a small output data set, making
use of large constant data structures."
Clearly, a programmable GPU ALU would be more efficient than using
general purpose CPUs.
--
JRT
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)
