On 11/21/2011 06:33 AM, Jayanta Roy wrote:
Hi Marcus,
The fft benchmark we have used are based on single precision 1d real
ftt running
on Intel Xeon platform. The ipp library version used was 5.2. The fftw
was of 3.1.2
version. We saw factor of 3 speed-up for larger fft size in case of
ipp. E
Hey Marcus,
One thing that has helped us a fair bit is keeping a close eye on
interrupts. In particular making sure that you set appropriate IRQ
affinities (via /proc/irq/irq#/smp_affinity), so that one core is not
doing all the
lifting when it comes to I/O and other tasks. In our case keeping the
On 11/20/2011 03:47 PM, John Ford wrote:
Have you done any profiling to see where the cycles are going? It seems
on the surface that your 6 core Phenom system should be capable of
processing 50 Megabytes per second.
John
It's not the data rate that's the issue, I can happily consume 50Msps
On 11/20/2011 12:57 PM, Yashwant Gupta wrote:
Hello Marcus,
At the GMRT, we have a fully real-time all CPU based correlator +
beamformer + pulsar receiver that takes in data from 64 analog inputs
(32 antennas, dual polarization) each at 66 Msps max.
I refer you to a publication that discuss
Hi Marcus,
I felt compelled to chime in. We have no direct experience with Casper
hardware - but have deep experience in both opencpi.org and netfpga.org .
In both cases we are in the business of allowing heterogeneous compute
resources (e.g. GPP, GPU, FPGA) to interoperate. At the risk of stating
>
> I'm involved a lot in Gnu Radio and the USRP hardware from Ettus, rather
> than the Roach boards from CASPER. But I have some questions
>about host-side processing, that the folks on this list might have
> some insights into.
>
> The USRP hardware was just upgraded to allow the use of 8-bi
Hello Marcus,
At the GMRT, we have a fully real-time all CPU based correlator +
beamformer + pulsar receiver that takes in data from 64 analog inputs (32
antennas, dual polarization) each at 66 Msps max.
I refer you to a publication that discusses the implementation in reasonable
detail :
R
On Sun, Nov 20, 2011 at 9:26 AM, Marcus D. Leech wrote:
> My current interest is CPU-only. Most of the papers I've read about
> GPU-based implementations aren't terribly encouraging for
> implementations where only part of the job is performed on the GPU. The
> overhead costs are killer, which
hi marcus,
the vegas software has an option for CPU only mode,
where all the processing is done on the CPU, and
the GPU isn't used.i think the CPU mode
can process about 100 MHz bandwidth, depending on
motherboard. see VEGAS docs on how to use
the software and set flag for CPU mode only.
On 11/20/2011 12:10 PM, Jason Manley wrote:
I suppose it depends on your definition of trivial! :)
Simon Ratcliffe (cc'd here) at KAT has a GPU-accellerated PFB working at
800Msps on 8 bit data using ~1 yr old hardware. Not sure how helpful this is if
you're not interested in using GPUs too. T
hi marcus,
there are several casper instruments that digitize data using
ADC's plugged into one or more FGPA board(s).
The FPGA boards packetize and time stamp the ADC data
and send it over 10Gbit ethernet
to one or more CPU's and GPU's.
here are a couple of examples.
the VEGAS instrument proc
I suppose it depends on your definition of trivial! :)
Simon Ratcliffe (cc'd here) at KAT has a GPU-accellerated PFB working at
800Msps on 8 bit data using ~1 yr old hardware. Not sure how helpful this is if
you're not interested in using GPUs too. There are a few pulsar guys doing neat
work at
Hi Marcus,
I wonder if anyone on this list has done more-than-trivial processing,
in real time, on sample streams approaching 50Msps in a host
computer environment, as opposed to the usual FPGA environment.
If so, what kind of CPU? How much memory? What kind of network interfaces?
Adam Deller
I'm involved a lot in Gnu Radio and the USRP hardware from Ettus, rather
than the Roach boards from CASPER. But I have some questions
about host-side processing, that the folks on this list might have
some insights into.
The USRP hardware was just upgraded to allow the use of 8-bit samples,
w
14 matches
Mail list logo
