Transcript Slide 1

Casper working group on
Hybrid Backends and Computing
ADC # 1
FPGA # 1
SWITCH
ADC # N
Compute
Engine # M
FPGA # N
CASPER meeting
Compute
Engine # 1
Capetown
1st October 2009
List of topics for discussions
 Specialised roles for different kinds of processing units : esp GPUs
What are GPUs best suited for ? Are there software platform issues with GPUs ?
Are GPUs just a “flavor of the season” phenomenon ?
 Interoperability : CPUs / GPUs in place of FPGA unit (e.g. as a test bed) : can we
have a few generic routines for this that can be made available by Casper ?
 Software correlators : can we have repository of tools and techniques at Casper ?
 Raw voltage recording and high speed data storage
 Going beyond the “real-time” processing : e.g. full analysis pipeline in real-time
(applications that require full floating point accuracy) – will PAPER lead the way ?
 Uniformity of programming tools and techniques :
 Languages
 Issues relating to “standard interface” : packet format – can we agree on this?
 Infrastructure issues : can we / should we compare the FPGA / CPUs / GPUs for
items like compute power per watt ; compute power per dollar/euro/rand/…;
compute power per man year of investment…
Suggested list of topics for discussions












Do we really need anything more than FPGAs ?!? -- no discussion needed.
What are the other options : CPUs, GPUs, other kinds of processors ? -- none.
Specialised roles for different kinds of processing units : esp GPUs
Interoperability : CPUs / GPUs in place of FPGA unit : can we have a few generic
routines for this?
CPUs and GPUs : how far can we go with these ?
Mix and Match of FPGAs and CPUs/GPUs : based on needs / functionalities :
 e.g. pulsar processing vs interferometry
Software backends for special applications :
 Spectral zoom modes
 Raw voltage recording modes
 Debugging mode for new arrays
Going beyond the “real-time” processing : e.g. full analysis pipeline in real-time
(applications that require full floating point accuracy)
Uniformity of programming tools and techniques : how crucial is this ?
Languages
Issues relating to a “standard interface” : packet format
Infrastructure issues : compute power per watt ; compute power per
dollar/euro/rand/…; compute power per man year of investment…
Hybrid Computing : SUMMARY
 GPUs
 have a useful role to play in the present scenario, in spite of some issues related to
programming language, not-yet-demonstrated capability for correlations.
 Groups will continue to use them in the short & mid term; Casper should welcome
and house the repository of GPU techniques for the benefit of the community.
 FPGAs + GPU systems : may be worth considering
 Software Backends :
 Generic term to cover software correlators, beamformers, pulsar processors, high
speed voltage recording (and playback) devices.
 Casper should have a repository for this on the website (Andrew S. is the contact)
 Going beyond the “real-time” processing : e.g. full analysis pipeline in real-time : looks
a bit early within Casper at present, but we should keep a look-out for how things are
evolving; also get inputs from other groups (e.g. MWA, LOFAR…)
 Issues relating to “standard interface” : packet format : Jason to put up the Casper
“philosophy” memo and solicit feedback
 Infrastructure issues :
consensus is that we should stay away from direct
comparisons of these kinds for the present (too complicated and unpredictable, could
discourage free exchange of ideas about the different technologies…)