[Open-graphics] CPU on graphics card]

[Timothy Normand Miller]

On 4/18/07, Hugh Fisher <hugh.fisher at anu.edu.au> wrote:
> Simon wrote:
>
> > Regardless of the cost of designing a better solution, my point is
> > that using a general purpose CPU is likely to be infeasible, because
> > the price will be too unattractive for the hardware to be profitable.
> >
>
> This discussion has now reached the point where we need some
> hard numbers from a hardware person (which I'm not). So who
> has experience on buying a CPU (ARM component or complete
> CPU) vs developing custom floating point ALUs? Timothy?

I can't give you hard numbers.  But something we can do is try to
calculate the kind of throughput from a CPU versus the throughput from
a hard-wired circuit.  Indeed, we already know that we can make 100MHz
logic in an FPGA outperform a CPU at very specialized tasks.  This
doesn't answer the question about what kind of throughput we'd get
from an array of SIMD processors vs. an array of CPUs.

We assume that the specialized processors would be faster.  But can we
prove it right now?

This is an analysis I'd love to do, but I'm going to have to put off
for a while.  Meanwhile, I can talk to some of my classmates who have
worked with "dataflow" architectures to see how they perform.  I'm
imagining our GPU design as a set of SIMD processors that feed other
SIMD processors in a pipeline.

-- 
Timothy Normand Miller
http://www.cse.ohio-state.edu/~millerti
Open Graphics Project