BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:More FLOPS or more precision? Accuracy Parameterizable Linear Equa
 tion Solvers for Model-Predictive Control - Antonio Roldao (PhD@IC)
DTSTART:20090205T120000Z
DTEND:20090205T130000Z
UID:TALK16096@talks.cam.ac.uk
CONTACT:Dr George A Constantinides
DESCRIPTION:In this work we exploit FPGA flexibility in the context\nof ac
 celerating the solution of many small systems of linear equations\, a prob
 lem central to model predictive control (MPC). Using iterative methods for
  solving these systems\, one can obtain an improved accuracy either by run
 ning for more iterations or by using more precise internal computations\, 
 unlike direct methods\, where accuracy is only a function of operation pre
 cision. Thus\, in iterative methods\, for\na given accuracy requirement we
  may conduct fewer iterations in a higher precision\, or more in a lower p
 recision. We argue that this suits FPGA architectures ideally\, as low pre
 cision operations result in greater parallelism for any fixed area constra
 int. We show that we may therefore optimize the performance by balancing i
 teration count and operation precision\, resulting in a several-fold speed
  improvement over a double-precision implementation\, but with the same fi
 nal result accuracy. Exploring this trade-off it is possible\nto provide a
  speed-up of 23 times on average\, 10 on the\nworst case and 38 on the bes
 t\, compared to a high-end CPU running at 3.0 GHz. This has the potential 
 to allow modern high-performance control techniques to be used in novel se
 ttings such as aircraft and diesel engines.\n\nJoint work with Amir Shahza
 d\, George Constantinides\, and Eric Kerrigan
LOCATION:Mahanakorn Laboratory\, EEE
END:VEVENT
END:VCALENDAR