arthur@uunet.UU.NET (Arthur Veen) (06/15/88)
Just to set the record straight: allthough the dataflow field is full of paper designs, a few machines have been built, tested and (to some extent) evaluated. In [1] I list 12 machines that I assumed to have reached the operational stage. I was too optimistic about Arvind's machine and I have not heard anything about the japanese PIM-D either. At the other hand the Sigma-1 is probably operational by now. That leaves 10 or 11 machines that have actually run programs. Of these the Image Pipelined processor uPD7281 from NEC is a static dataflow processing element on a chip that has been incorporated in several machines (of which I have no details). Finally , work is going on here in The Netherlands on a commercial dataflow machine of which the first prototype is nearing completion. Keep a look-out for its announcement. The dataflow field is less fashionable (and less crowded) than it was 5 years ago but certainly not dead. [1] A.H. Veen "Dataflow Machine Architecture" ACM Computing Surveys 18(4), Dec 1986 Arthur H. Veen Parallel Computing Postbus 16775 1001 RG Amsterdam The Netherlands
ian@esl.UUCP (Ian Kaplan) (06/17/88)
In article <1919@hubcap.UUCP> mcvax!nikhefh.hep.nl!arthur@uunet.UU.NET (Arthur Veen) writes: >Just to set the record straight: allthough the dataflow field is full of >paper designs, a few machines have been built, tested and (to some extent) >evaluated. [ text deleted ] >Sigma-1 is probably operational by now. That leaves 10 or 11 >machines that have actually run programs. Of these the Image Pipelined >processor uPD7281 from NEC is a static dataflow processing element on a >chip that has been incorporated in several machines (of which I >have no details). >Finally , work is going on here in The Netherlands on a >commercial dataflow machine of which the first prototype is nearing >completion. Keep a look-out for its announcement. >The dataflow field is less fashionable (and less crowded) than it was 5 years >ago but certainly not dead. The Japanese Sigma is indeed completed. They have something like a 128 processor machine. They also have some kind of data flow compiler, but I have been unable to find any details. I hope that the Sigma group will have some papers in this years International Conference on Parallel Processing. Currently, the Japanese seem to be the leaders in data flow research. As to the NEC uPD7281, it does indeed use data flow principles, but it is difficult to build a system with a lot of processors, since the chip was not designed for this. Also, the chip does not have any "hooks" for implementing anything like demand driven dataflow, which means that it is not practical for implementing a general purpose data flow machine. I hope that if any details on the Netherland machine become available they will be posted on the net. Data flow may not be dead, but it sure seems to be in a deep sleep. This sort of reminds me of a Monte Python skit about a Parrot.... If anyone has details on the Sigma machine, I would love to see those as well. Ian L. Kaplan ESL, Advanced Technology Systems M/S 302 495 Java Dr. P.O. Box 3510 Sunnyvale, CA 94088-3510 ian@esl.COM esl!ian@ames.arc.nasa.gov