Slashdot Mirror
DIY 1980s "Non-Von" Supercomputer
Brietech writes "Ever wanted to own your own supercomputer? This guy recreated a 31-processor SIMD supercomputer from the early 1980s called the 'Non-Von 1' in an FPGA. It uses a 'Non-Von Neumann' architecture, and was intended for extremely fast database searches and artificial intelligence applications. Full-scale models were intended to have more than a million processors. It's a cool project for those interested in 'alternative' computer architectures, and yes, full source code (Verilog) is available, along with a python library to program it with." Hope the WIPO patent has expired.
Check this out...
/dev/null.
Article at
Everything that you wanted to know and more. An interesting read.
Wasn't the transputer an example of this architecture? I'm old enough to be able to say "Get off my lawn!" and remember when the transputer came out; it caused quite a stir.
Tiller's Rule: Never use a word in written form that you've only heard and never read. You will end up looking foolish.
Folks just don't understand what FPGA's do.
"So, that's neat and all but did I misunderstand something. His model doesn't seem that powerful unless he was using modern processors?"
It's implemented in HARDWARE. Everything runs in parallel. To do the same on a "modern" processor, would require 300-400Mhz. A FPGA running at a [modest] 25Mhz could get the same or better performance.
"I can't help but wonder if this couldn't be emulated for a fraction of the price. Are there any virtualization systems out there that could accomplish what this guy did? I imagine something along the lines of GNS3 might work..."
FPGA's are cheap. A Spartan-3 board can be had for 100-200, and probably hold 2-3 32 node cpu's.
Programmers just don't understand the difference between say verilog, and C/C++/Java.
verilog is the basic building block of CPU's. Everything is done in PARALLEL by default, while in C++/Java everything is done SERIALLY.
Wasn't the transputer an example of this architecture? I'm old enough to be able to say "Get off my lawn!" and remember when the transputer came out; it caused quite a stir.
The transputer was a RISC-ish CPU with 4 high speed DMA/serial links allowing it to be easily connected to other Transputers (each with its own local memory) to form a network. As such, it could be used to build a large MIMD system - not a SIMD one.
Transputers (+ the Occam language) supported multi-threaded programming with very fast context switches and, for its time, they also had very good FP performance when compared to the contemporary x86+float coprocessor.
Of course a modern computer can simulate a 1980's computer. It would probably take about a day to write a functional simulation in Java.
For that matter, it's not like this computer can do anything that a modern computer can't do in spite of the different architecture. It was designed to do certain things fast, but anything off the shelf today could run circles around these relics regardless of such optimization. (To GP's point -- since the article indicates that he was building to the functional design of the original, it's probably not powerful by today's standards. He may have used faster components than they had back then -- and he obviously used smaller components than they had back then -- but we're not looking at a modern billions-and-billions-of-transistors-on-a-chip optimized-in-ways-you-cannot-comprehend heat-sink-needing CPU.) So once you talk about "what can it do" at a useful level of abstraction, the answer is "nothing all that practical".
But that's not the point, is it? This kind of stuff is a hobby and a fascination to some people. I'm interested enough that I might write a software simulation of the machine, but not interested enough to build one. This guy was interested enough to build one.
It's not like stamp collectors are saving up for a big letter-writing campaign...
For several years I worked on a SIMD system called MasPar. The system had 8192 processors. It was installed in 1991 and it was not until about 1998 that conventional computers running Oracle could even come close to the performance for data warehouse applications. Sure, it's slow by today's standards, but I bet a modern version custom built would be an awesome code breaking and data analysis system.
BTW: the system was used to help with the human genome project and to search Medical Services Plan data by the Province of BC. It finally decommissioned in 2000 (or early 2001).
Uhhh... right here. Now don't say I didn't warn you, but it's pretty relevant.
Link attempts to install a virus; need to kill his comment.
I got a chance to use a Connection Machine (real, not emulated) in the late 1980s, just a couple of years out of college. It was an internal R&D project for a defense contractor, porting a computational fluid dynamics program I didn't understand from Cray vectorized Fortran to the CM's *Lisp. Fun stuff.
I even got a chance to visit Thinking Machines headquarters in Boston, and hear Danny Hillis speak. Here he was speaking to a room full of suits, dressed in jeans, sneakers, and a T-shirt. I remember thinking at the time that being able to do that was quite an indicator of success.
Yeah, yeah, I know...offtopic, overrated, etc. So mod me down if you must. (Or is that just reverse psychology on you moderators? Muhahaha!)
"Rub her feet." -- L.L.