New Supercomputer By Star Bridge

Ronin Developer writes with word of this "interesting article on CNN about a new desktop-size super computer that reconfigures itself on the fly. The company name is 'Star Bridge.' Ring any bells? If I remember correctly, wasn't there something on /. about this a year ago?" Indeedy do -- Star Bridge seems to go straight from wacky-but-cool promises to Where are they now? (and back) with finesse. It's the the Arnold Schwarzenegger movie plot of hardware companies -- simultaneously head-scratchingly implausible, mildly compelling, and numbingly persistent.

Comments

The Real Issue

[deadline]

Some reality is in order here.

FPGA computing is real and it has been shown to work for some problems. Take a look at TimeLogic. These guys have implemented search algorithms used in the human genome project on FPGAs.

Now let's look the difference between "works" and "price to performance". In the case of TimeLogic they have produced a "stand-alone appliance that end-users do not program (i.e. users do not program the FPGAs.) I beleive the reasons for this is that this programming abstraction (remember this) is not easy to master (i.e. it is not a mainstream programming language.) Nor is the "edit, complie, run cycle" easily reporduced on a desk top. (This time is perhaps the single most limiting factor in software production). So FPGA computing works, but is expensive to implement and program. It does not support cost effective general programing practices that are used today.(i.e. unless you are building specific purpose machine and can justify the software development costs based on a real market, the cost of programming for every day production environments is too expensive)

Which brings me to the main point. The issue is SOFTWARE. It is easy to build a Beowulf with 1000 processors and call it a supercomputer. It is hard (expensive) to write good software for this system. It is easy to string together a bunch of FPGAs and call it a supercomputer. It is hard (expensive) to write software for these things and it is harder (expensive squared) to write parallel software for these things.

In general, there is a huge (I mean really huge) investment in the supercomputer world in programming abstratcions that use FORTRAN (and to some extent C) Side Note:Before all you "FORTRAN is dead language boneheads" start hitting the reply button, remember that there are more than a few 100,000+ line FORTRAN programs that determine everything from airplane wings, to weather, to new drugs, that are not going to go away because you think XML is great way to go. Indeed, the cost of reprogramming these applications is almost an economic impossibility!

So where were we, ah yes, the software thing. My point is that until FPGA systems can take standard supercomputing FORTRAN or C applications and run them "out of the box" and thereby allow the tens of thousands of people who understand this type of programming to use FPGAs easily, they will remain application specific computers (albeit fast) and not realy a mainstream programmable computing devices. This is not to say in the future the FPGA computing will not dominiate (maybe it will), but there is a lot of work to be done on the software side before this will happen.

BTW: I sent the Starbridge guys some simple FORTRAN benchmarks a while ago. I did not receive a response.

Finally, remember this:

The general always eats the specific.

Any one remember a company called Symbolics?

FPGA's

[Grant Elliott]

These things use Field Programmable Gate Arrays (FPGA's) in order to restructure themselves dynamically. This, in and of itself, is not a new concept. FPGA's have been used for years in prototyping or in the first products released. It's much cheaper and easier to reprogram an FPGA if a bug is found than it is to create a new chip design. Once the bugs are gone, FPGA's are replaced by hard-wired silicon in the rest of the line.

Now on to using FPGA's in supercomputers. First of all, an FPGA is slower than a hard-wired chip. These machines pick up speed from the fact that they can use portions of the chip that otherwise would have been on standby. It's super-charged serialization. By restructuring the circuitry for each task, they can take advantage of the majority of the chip at all times. This is not an easy task, and I find it quite impressive. (On a side note, the restructuring is software-controlled.)

When I read this story, I immediately associated it with an article from several years back about Inman Harvey and Adrian Thompson. Thompson was using an FPGA to run genetic algorithms for hardware development. Essentially, make a machine design the chip. He had some very interesting results. The chip designs took advantage of the physical chip rather than just the wiring. They were incredibly efficient, but Thompson couldn't understand why they worked. (He suspected such things as electromagnetic coupling and communication through the power supply.) This is all only moderately related, but it's very interesting, regardless. The article is from June, 1998 and can be found here if anyone is interested.