SLIDE 1
Application specific processors for Application specific processors for intensive computing applications intensive computing applications University of Pavia, Italy Faculty of Engineering Microcomputer lab
Gianni Danese – Full Professor Francesco Leporati – Associate Professor Mauro Giachero – Ph.D. Student Nelson Nazzicari – Ph.D. Student
SLIDE 2 Overview of our activities
FPGA for high performance computing
Montecarlo simulation Fingerprint matching
Real time video compression of Formula 1 sensor
data
Motor control for wind tunnel equipment Intelligent cameras
SLIDE 3 FPGA for high performance computing (1/2)
Dipole cubic lattice model – Energy
calculation
Montecarlo simulations need a lot
SLIDE 4 a lot of repeated, regular
calculation
high-performance VLIW style architecture
FPGA for high performance computing (1/2)
SLIDE 5
FPGA for high performance computing (1/2)
Our current architecture features a speedup of 2
compared to a 3GHz Pentium 4 processor
The 1S40 FPGA can host 2 accelerator, bringing
the speedup factor to 4
Using the more recent Stratix II or Stratix III the
achievable speedup can be even greater
This way is possible to build up an accelerator
cluster whose cost/performance ratio is much smaller than that obtainable with a cluster of PCs
SLIDE 6
Fingerprint matching need a lot of computation
FPGA for high performance computing (2/2)
one new image vs. a
big archive
SLIDE 7
(band limited) phase only correlation algorithm
FPGA for high performance computing (2/2)
correlation phase-only correlation
SLIDE 8
A little spin-off: high performance DFT on FPGA
commercial solutions are not satisfying (too
slow, big or inaccurate)
butterfly algorithm (created for parallel
implementations)
different attempts (fixed vs floating point, data
size, etc.)
better than ALTERA FFT Core and other
commercial products (using our application- specific metric)
SLIDE 9
Video Compression on F1 vehicle
thermic camera to control
tyre temperature
normal b/w camera to
control tyre deformation
low band high noise
channel between car and boxes
SLIDE 10
Video Compression on F1 vehicle
need for real time on
board video compression to decrease the amount of data to be transmitted
some on board data
elaboration
SLIDE 11
Temperature evolution of four tyres as seen by thermic cameras Image and video acquisition/compression platform (with extra debug/develop features)
Video Compression on F1 vehicle
SLIDE 12
actuator control for 1:2
model placement
position and speed sensors treadmill rolling control closed loop control to
maintain belt alignment Motor control for wind tunnel equipment
SLIDE 13
VISoc: Intelligent vision system
intelligent camera with
embedded neural network
object recognition shape count
(cars, persons...)
SLIDE 14
Partnerships
Magneti Marelli Ferrari Neuricam Assomac ST Microelectronics National Institute for the Physics of Matter National Institute of Nuclear Physics Polytechnic of Milan