On 22/06/2011 10:27, RussellMc wrote:
>> You could probabblly optimise quite a bit over a generic SHA256
>> >  implementation but I still think you would struggle to keep up with G=
PUs
>> >  using reasonablly priced FPGAs.
> Why?
> What if anything does this say about the ability to use a GPU to carry
> out tasks usually done by FPGAs?
>

GPU vs FPGA performance per $/ Euro/Yen/Bitcoin  depends on the kind of=20
problem solved/task implemented. A very wide range of tasks done by FPGA=20
are impossible on GPU. Some tasks GPU better. Some applications an 18F=20
PIC beats both!

GPUs are very limited to parallel structured repetitive computational=20
tasks (not sure how well pipelined is done). A PIC 18F may be superior=20
for a slow complex purely sequential program with many decisions and=20
little or slow Floating point.

An FPGA is poor for stored program computation tasks, emulating a GPU or=20
CPU, though allows fast prototyping of new ones. It performs best with=20
tasks that can be "hard wired" with no program instruction execution and=20
essentially repetitive Integer multiply / accumulate of word width equal=20
or less to physical MAC units. One FPGA has maybe 1/10th to 1/20th=20
number of Integer MACs compared to number of cores able to do vector or=20
FP arithmetic in a single GPU. Low end FPGA is very much cheaper than=20
GPU. High end FPGA is very very much more expensive than GPU chip.

I'd not like to guess which is more cost effective for SAH256 at same=20
speed. Or which can go faster on one PCIe plug in card if money no=20
object. You can get high End FPGA on PCIe card. So no HW design required.

Obviously to prototype an ASIC with DSP, or high speed interface between=20
"things" even a cheap FPGA  beats PIC18F, Xeon or GPU.

--=20
http://www.piclist.com PIC/SX FAQ & list archive
View/change your membership options at
http://mailman.mit.edu/mailman/listinfo/piclist
.