Lambda the Ultimate

Architecture is bloody important. There's all sorts of neat stuff languages and operating systems (especially operating systems) could do if the architectures but allowed it instead of just pushing constantly for a faster x86.

It wasn't very clear from the article, but you can get it at http://beecube.com/

That is an excellent article. It brought back a lot of feelings regarding my former foray into computer architecture research. The most relevant points are that a lot of the research didn't have any future and that much of it was not applicable to real-world situations. The latter is not so bad by itself, but along with the former meant that it would not eventually become applicable.

I certainly hope they contribute some revitalization to the field of computer architecture research. It needs it.

Thanks for post, Luke!

“If you want to try out a new architecture or a new feature,” Thacker says, “it’s very difficult to do these days, because you have to actually build chips to build anything of substantial size. Using FPGAs, you don’t have to build chips..."

...

The viability of the FPGA approach was demonstrated last summer, when Zhangxi Tan, then an intern at Microsoft Research Silicon Valley, built a system for solving the problem of binary satisfiability, commonly used in design automation.

“He got much faster speed than what a computer could do,” Thacker reports, “because the algorithm is exactly suited for what can be done in FPGAs.”

I was involved with one 'cute' CPU arch research, and I got out of the field precisely for the first quote above. The thing with engineering research is that it's firstly about engineering, and you don't get a prize for the cutest block diagrams if the company with the dropped e in its name has chips that are better than yours, because engineering is about things like a huge corporate budget, voltage levels, crosstalk, processing technology, manufacturing, C compilers and drivers for power management as much as cute block diagrams drawn by a few people. Playing with FPGAs (for general purpose CPU research, not application specific stuff like the second quote) is just pretending.

By "current machines", I mean off-the-shelf computer systems which a reasonably well-funded research organization can afford.

Certainly, there are ways in which off-the-shelf systems are sub-optimal. But the situation we have today is a lot less dire than the situation facing Alan Kay and his team at PARC, were the off-the-shelf computing technology of the 70s (ugly terminals if that, mainframes/minis) were inappropriate for what they were trying to build. (And Smalltalk is an unusual case because of the whole environment; an OO-based Lisp derivative with unusual syntax sans the Smalltalk environment certainly could have been hosted on a terminal of the time).

The Lisp Machine may be a more interesting case of an alternate computer architecture (especially down at the micro-architecture level).

But what cannot be done--or is only done with great difficulty--on current machines? The bottleneck on most modern computers these days isn't CPU speed, it's I/O latency and bandwidth. Parallelism is one way to attack the bandwidth limitation, and FP is a good way to make parallel computation more tractable; but existing computer architectures can be used to enable PL research here.

By "current machines", I mean off-the-shelf computer systems which a reasonably well-funded research organization can afford.

Like RAMP?

"... little is known on how to build, program, or manage systems of 64 to 1024 processors, and the computer architecture community lacks the basic infrastructure tools required to carry out this research."

We just bought one of these at Sun Labs and it's pretty sweet. It makes a hell of a lot of noise and heat, though. It has its own office now ;)