e122-cl@polonius.berkeley.edu (Timothy Kong) (12/15/88)
Simple question: What are transputers? Someone mentioned them in a few articles back but never explained what exactly they are. I have heard talks about transputers' incredible processing power and how adding one of them would drastically speed up your machine. Could somebody please elaborate.
jbn@glacier.STANFORD.EDU (John B. Nagle) (12/15/88)
In article <8304@pasteur.Berkeley.EDU> e122-cl@polonius.berkeley.edu (Timothy Kong) writes: >Simple question: What are transputers? >Someone mentioned them in a few articles back but never explained what exactly >they are. I have heard talks about transputers' incredible processing power >and how adding one of them would drastically speed up your machine. >Could somebody please elaborate. The INMOS Transputer is a microprocessor from INMOS, an English computer firm, designed especially for use as a component of parallel computers. One 32-bit transputer is roughly comparable to a M68020 (+- 50%). But they are intended for use in large arrays, perhaps as many as 1024 processors. The machines have special hardware to support inter-CPU communication in such arrays. Adding one of them is not particularly useful, but adding many of them makes a supercomputer of sorts. Add-on transputer boards (typically with 4 transputers to a board) are available for the PC/AT and Mac II. Larger arrays are available for workstation-class machines. INMOS offers an interesting language called OCCAM for programming transputer arrays, but a modified C seems to be replacing OCCAM. In either case, one must deal with concurrency explictly. The model is a non-shared memory MIMD machine, with relatively large memory, typically 64KB to 1MB, per processor. The transputer has been out for a few years now, and has a reputation of being a solution looking for a problem. If you have a problem that slices up well into parallel transputer-sized units, it's a very effective way to go. But dividing up problems for it is difficult, OCCAM notwithstanding. So it tends to be used as a component of special-purpose machines, like CT scanners, signal processing systems, and the like, rather than as a general-purpose computing element. This is the typical curse of moderately-parallel machines. See comp.sys.transputer for more info. John Nagle
grabas@m.cs.uiuc.edu (12/16/88)
Talking of transputers in a few words is hard, and I have very little time (Finals' week...). To have more information, you could read the Transputer notesfile. I don't know how it si called at berkeley; here (U of I) this notesfile is called comp.sys.next and the transputer one is called comp.sys.transputer. A transputer is an integrated circuit comprising: -A 32-bit RISC-like CPU -A 64-bits FPU (IEEE standard) -A certain amount of on-chip memory (4K currently) -A certain number of communication links for inter-transputers communication. Each link implements two DMA channels (one read, one write) between memory and a bit-serial line. the links work at up to 20Mbits/ second; there are 4 of them on the current implementation (T800) -A memory controller which allows the connection of a transputer do dynamic RAM without any additionnal logic (the controller takes care of the refresh cycles, etc...) To sum-up, you can built a parallel machine using transputers connected one to the other with their links; you do not need any memory with your transputers because they have on-chip memory and can be booted from a link instead of from ROM. Anyway, if you want to boot them from ROM, or increase their memory using external memory chip, you do not need any extra logic (well, may be an address latch, because addresses and data are multiplexed on certain models) to run the system. This may not be 100% clear and detailed, but I do not have a lot of time today... Dominique Grabas, University of Illinois at U-C.