Slashdot Mirror
Self-wiring Supercomputer
redcone writes "New Scientist is reporting on an experimental supercomputer made from Field Programmable Gate Arrays (FPGA) that can reconfigure itself to tackle different software problems. It is being built by researchers in Scotland. The Edinburgh system will be up to 100 times more energy efficient than a conventional supercomputer of equivalent computing power. The 64-node FPGA machine will also need only as much space as four conventional PCs, while a normal 1 teraflop supercomputer would fill a room. Disclaimer: At this point in time, the software needed to run it, which is the key to the project, is vaporware. "
How does the playstation 3 manage 2.2 teraflops without being the size of a house then?
It's been a while since I learned about this stuff, but hasn't it been mathematically proven that it's futile to try to write a program that, presented with an arbitrary (expressable) problem, will write a program to solve that problem?
As already mentioned, the biggest problem with FPGAs is the difficulty/time in writing the logic. While that's not necessarily a big problem for a major supercomputing center or a CS research center, it (along with cost) is a problem that prevents FPGAs from being routinely adopted by end-users such as people in the applied research community.
One idea to get around this has been advanced by (among others), Stretch, Inc.. The summary is that their compiler analyze your C-code and decide what can be more effectively rewritten as new instructions for their chip, and sets it up on the fly. You never get the ultra-low level control (or performance) of FPGA programming, but in principle you get more performance than before.
Their primary applications have been basically as programmable replacements for DSPs, but they really want to push workstation applications for their products.
That being said, I neither work for them nor have I ever used any of their products, but it certainly sounds interesting!
Although using FPGAs for reconfigurable computing applications still has a number of drawbacks, utilizing FPGAs for embedded applications is some really cool stuff.
For example, an entire system can be dynamically built right into the FPGA -- including processor, OPB, memory buses, and any other devices such as interrupt controllers, timers, etc. Aside from RAM and Flash, you almost have an entire embedded system built right into a chip.
Earlier this spring I had the opportunity to work on a project that required this very embedded setup. Using the MicroBlaze soft-processor from Xilinx built into the Spartan3 FPGA and only 8MB of SDRAM, I got uCLinux running -- completely tailored to my hardware setup!
I can't tell you how much time and money would have been wasted trying to design and fabricate the same setup on a PCB.
GenoByte has found a more novel use for FPGAs, which they call "evolvable hardware." Much like our own brains neural networks on the FPGAs reconfigure the way they interconnect on the fly; commonly used paths are reenforced while less frequently used ones atrophy.
Here are some cool pictures:
The CAM-BRAIN machine, a big box full of FPGA boards: http://www.genobyte.com/images/machine.jpg
Neural network layout for the XC6216 FPGA: http://www.genobyte.com/images/chip.JPG
All in all this approach is substantially faster than modelling large neural networks on a general purpose processor. In the GenoByte approach, the neural network is implemented as physical circuits.