Re: [Discuss-gnuradio] PC performancee

[Eric Blossom]

On Sat, Jan 11, 2003 at 09:34:49PM -0800, John Turner wrote:
> I made a general assumption that the host PC has limited processing
> capabilities, i.e. limiting the maximum effective sampling rate of 500 to
> 1MSPS. I'm curious, what can a PC realize in terms of processing benchmarks
> at sampling rates of up to 20MSPS? And are there typical formats that are
> desirable FM/AM/SSB, IEEE 802.11, Bluetooth, GPS, OFDM, CDMA.

Our current code, doing everything in software, with a 20MSPS sampling
rate, can demod 2 broadcast FM signals, using floating point.  We
should, with about a day or two's work, be able to do 4 channels if we
use fixed point (MMX).  This is using either a 1.7 GHz Pentium or an
AMD Athlon 1800+ (1.5 GHz clock).

A bit of current reality: our SIMD floating point FIR code comes very
close to 1 cycle / tap on both the P4 and the Athlon.  This gives
about two billion taps / second on contemporary commodity hardware.

Note that by far the vast majority of cycles in narrow band
applications are burned in the software channel selection code.  This
is the first thing that would be put into the FPGA (digital
downconverter). Hence, I wouldn't be surprised to find that we could
do 30 or 40 channels of FM on the same platform, concurrently.  I've
heard estimates on the order of 20k - 30k gates for the digital
downconverter in the FPGA.  This includes the NCO and CIC filters.
There are of course trade offs you can make between how much
channelization is done in the FPGA (limited by gate resources, and bus
speed) and how much you do in software (limited primarily by MIPS).

Eric