View Full Version : 3 PS3 = 140 times faster than C2D 2Ghz on a brain algorithm for vision and language
Playstation3 helps robots see
Rick Merritt EE Times (09/24/2007 7:59 HŒ EDT)
AUSTIN, Texas — Robots took a tiny step closer to seeing the same way humans do thanks to a team of university researchers who ported to the Cell processor new vision algorithms derived from brain research. The team from Dartmouth and the University of California at Irvine were able to get three networked Playstation3 devices to recognize a given object in one second using their software.
"We aim to put all the speech and vision recognition into a working robot, so we need real-time performance," said Felch. "DARPA wants to see people create robots that can actually drive a vehicle without harming anyone in the process," he added.
The team spent about eight months on the project first implementing the algorithm on a 2 GHz Intel Core 2 Dupo processor. Using the PC, the team showed machine vision that could recognize in three minutes a bar stool in an image of an office setting.
Using a network of three Playstation3 consoles linked to a PC, the tem was able to speed the recognition rate up to just one second. "A one-second delay is essentially real time object recognition, and that is just what humans do," said Felch.
Thanks to its on-board accelerators, the Cell processor in the consoles was able to handle key computations in three cycles that the Intel chip had to compute sequentially in 15 cycles. Overall the three consoles handled the work at rates up to 140 times the speed of the single PC processor, Felch said.
The underlying algorithm breaks objects down into a hierarchy of key shapes and objects called line triplets. Those primitives are then compared to similar shapes in a new image. The research effort was focused on speeding up the process of making those comparisons.
Latency is significantly less important for the brain algorithm. Synapses typically have latency delays of one millisecond, Felch said. That's more than an order of magnitude higher than the latencies in today's fastest computers, he added.
Cool stuff! There's details of other winners here too:
Today at the 2007 Power Architecture Developer Conference (PADC), IBM (NYSE: IBM) announced the winners of its first annual Cell Broadband Engine(TM) (Cell/B.E.) Processor University Challenge. From the thousands of innovative entries, winning designs featured never-before-seen uses of the Cell/B.E. technology, including large-scale modeling of the human brain; a system for mapping massive amounts of real-time data; a path to deliver complex, 3-D medical images to a desktop computer; and a new way to detect extremely fast-moving objects.
Nearly 80,000 students from 25 countries competed in the Challenge, which consisted of online trivia about Cell/B.E. -- originally designed by IBM, Sony Group and Toshiba Corp., for use in consumer devices such as Sony Computer Entertainment's PLAYSTATION©3 -- followed by an opportunity to invent their own applications using this powerful processor. Students' designs included everything from applications-oriented solutions (e.g., visualization, medical imaging, seismic computing, etc.) to High Performance Computing (HPC) to industry-wide programmability tools.
"This contest provided a growth opportunity for students to gain real-life, multi-disciplinary skills to apply to their futures as they move from the classroom to the workforce," said Nick Donofrio, IBM executive vice president, Innovation and Technology. "This challenge also proved the true power, potential and promise of student innovations."
The teams with winning designs were each presented a cash prize ranging from $2,500 to $10,000 for their work. These included:
-- First Place -- Cluster of Sony PlayStation3's used for large-scale
modeling of the human brain. Using the same technology that runs in today's
video games, students Jayram Moorkanikara Nageswaran and Jeff Furlong from
the University of California, Irvine (USA), and Ashok Chandrashekar and
Andrew Felch from the Neukom Institute for Computational Science at
Dartmouth College (USA), created a low-cost cluster able to support the
complex algorithms used in brain research. This study addressed issues of
known difficulty in visual processing; for example, using standard
processors, the complex computations needed to emulate the human brain's
ability to rapidly and effortlessly recognize objects, was found to be slow
and inefficient. By exploiting Cell/B.E.'s parallel instruction set and
extending it into low-cost clusters using Sony PS3s, the students were able
to show a 100x performance boost over smaller clusters.
-- Second Place -- A new path developed for mapping large-scale data.
MapReduce for Cell/B.E. is a simple and flexible parallel programming
model, initially proposed by Google, for large-scale data processing in a
distributed computing environment. This implementation for Cell/B.E.
enabled programmers to easily use the resources of a large distributed
system. In a performance evaluation at the University of Wisconsin, Madison
(USA), student Marc de Kruijf used synthetic benchmarks representative of a
diverse application space. For computationally intensive applications, he
showed in excess of a 2.5x performance improvement over a 2.4GHz Intel
Core2 processor, with linear scaling as more Synergistic Processing
Elements (SPEs) were added. The runtime overhead was also minimal, at less
than 4 percent. This was the first application of its kind for Cell/B.E.
-- Third Place -- Complex 3-D imaging brought from devices, such as MRIs,
to the desktop. The importance of volume rendering has been increased as
the amount of data grows due to widespread use of 3-D imaging devices such
as Computed Tomography (CT), 3-D laser scanners and Magnetic Resonance
Imaging (MRI) equipment. The technique, called ray-casting, recognized as
one of the best for image quality, has been limited to a set amount of data
due to its slowness. The recent Cell/B.E. architecture provided
opportunities to finally put the ray-casting into the practical use at the
desktop computers of scientists and engineers. Jusub Kim from University of
Maryland at College Park (USA) presented a new volume ray-casting algorithm
designed to fully take advantage of Cell/B.E benefits and showed Cell/B.E
is the main enabling technology in providing the-finest-image-quality
volume rendering on practical data size. Experimental results showed one
can interactively render 256x256x256 data onto a 256x256 image at $\sim$15
frames/sec with one Cell/B.E processor, which was about 100 times faster
than the same implementation at Intel Xeon 3GHz.
-- Fourth Place -- A new way developed to detect fast-moving objects. A
project by students Robert Hiramatsu and Jussara Kofuji at the University
of São Paulo (Brazil) re-implemented rapid object detection on an Open
Computer Visual library (OpenCV) and used efficient ways to process on the
SPEs of CELL/B.E. OpenCV has direct relevance to cutting-edge visualization
applications such as facial recognition. In the team's implementation, they
used a specific approach of classifiers that restricted use of an image
reference of 24 x 24 pixels and worked with a stump-based classifier
algorithm to reduce data structure for classifiers.
In addition to the winners from the regions of North America and Latin America, students also participated from Europe and Asia. These winners will be recognized in a public event later this year.
Also gives an interesting perspective on how hard it is. Don't expect too advanced routines in games for this using the EyeToy or Vision cams on the consoles ... the whole thing about being able to identify a tennis racket in realtime and use that instead of a wiimote suddenly seems a little bit further off than some head-honchos suggested. ;)
I also like the implication that you can really just shell out 1500 bucks, get three playstations and do this kind of stuff.
That really is quite amazing. Is it down to that specific type of algorithm being better optimised for the cell style architecture or is it just raw performance, because how would a C2Q do in comparison to the dual-core results?
Cool stuff! There's details of other winners here too:
The contest is divided into 2 regions. CNN lists the winners in region 1 (Americas).
According to http://www-304.ibm.com/jct09002c/university/students/contests/cell/, the winning projects for region 2 (Europe and Asia) are:
* Exact CT Reconstruction on Cell Broadband Engine Architecture
Zhenxing Han, Yannan Jin, Liji Cao, Qing He, Si Chen and Xueming Yu Shanghai Jiaotong University
* Multi-resolution Texture Synthesis on Cell/B.E.
Wei Tong, Shien Deng and Gang Pan
School of Computer Science and Technology, Tianjin University
* H264 Real Time encoding on Cell/B.E.
Yao Zou, Xun He, Xianmin Chen and Lei Zhu
Shanghai Jiaotong University
* A Novel Grid Space with Cell Powered
Qi Lv, Chunfeng Si and Tao Yan
Also gives an interesting perspective on how hard it is. Don't expect too advanced routines in games for this using the EyeToy or Vision cams on the consoles ... the whole thing about being able to identify a tennis racket in realtime and use that instead of a wiimote suddenly seems a little bit further off than some head-honchos suggested.There's a big difference between recognizing an arbitrary object using a brain algorithm and following the motions of a known object. Following a tennis racquet is no different in principle to following the markings on an Eye Of Judgement card. The idea (I think it was from MS) of having a game recognize whatever you pick up is a long way off, but comparable Wiimote like activities aren't beyond the processing abilities of Cell.
vBulletin® v3.8.6, Copyright ©2000-2013, Jelsoft Enterprises Ltd.