A CPU is a device used almost entirely to process mispredicted
branches. Only when not tied up with this critical task does anything
else get done.

A GPU does not know what a mispredicted branch is, so 100% of its time
is available to perform computations. Because it has _many_ cores, it
can perform a great many computations in a short period of time, using
single precision FP and, perhaps, integers.

Robert Myers tells us the future of computing is in GPUs.

Question: is the Cell a CPU or a GPU? ;-)

On Tue, 22 Mar 2005 15:44:13 GMT, "Felger Carbon" <(E-Mail Removed)>
wrote:

>A CPU is a device used almost entirely to process mispredicted
>branches. Only when not tied up with this critical task does anything
>else get done.
>
>A GPU does not know what a mispredicted branch is, so 100% of its time
>is available to perform computations. Because it has _many_ cores, it
>can perform a great many computations in a short period of time, using
>single precision FP and, perhaps, integers.
>
>Robert Myers tells us the future of computing is in GPUs.
>
>Question: is the Cell a CPU or a GPU? ;-)
>
The PPC core is in order, and the SPE's sound like pure stream
processors (or DSP's). Cell is certainly geared to predictable
workloads. Call it whatever you like.

There _is_ a way to get a stream processor to handle branches, which
is to do both. The SPE's have dual datapaths, just like the C6000 DSP
series. Whether that programming technique is actually used is
another question. It certainly was on my favorite stream processor,
the Cray 1, where the vector mask register was responsible for picking
which of two results to use as the final answer.

As always, Felger, I participate in these forums more to be educated
than to educate, so I look forward to being educated.

RM