casey@lll-crg.llnl.gov (Casey Leedom) (04/21/89)
This is an issue which has bothered me for a long time. On the one hand we have a lot of manufacturers developing proprietary busses for their various systems, and on the other we do have several standard general purpose computer busses, but they're just too slow for the current and next generation(s) of systems being developed. I don't like the proprietary busses because they lock customers into the manufacturers' whims of production and pricing schedule, causing increased prices and diluted product availability. But to slave the new systems to the old bus standards is senseless. Why aren't manufactures getting together to define a new general purpose bus standard, or even a couple of them (ala memory, peripheral, special control, hierarchal and chained)? It really wouldn't take that much effort. The various manufacturers of new systems are having to define new busses in any case. All it would take is for a few of these manufacturers to get together and hash out the details of what's needed for the next generation standard bus. There are two possible followups to this letter: one on the technical problems of defining such a new bus and the other on the political. I'll prime both subjects with a few thoughts: TECHNICAL: [I hate news articles which start this way and you have every right to flame me in public for this: My background in hardware is almost non-existent. Therefore the following list may contain absurdities and obvious omissions. This is just to get the ball rolling.] Any new bus standard should probably have at least the following features: o Multiple busses/data-paths: memory, I/O, multi-cpu control and communication. o The bus system should be capable of hierarchal/chained setup. That is, you should be able to have multiple memory, I/O, etc. busses as needed by the system being put together. o Bandwidth on memory and I/O busses should be in the hundreds of megabytes per second. o Data and address lines should NOT be multiplexed. o Data and address paths should be very wide (probably 64 bits). o The bus needs to be simple enough that it doesn't make too many assumptions about system configuration. This will also make it possible to design a cheap VLSI bus controller chip that can be used to make it possible to design low cost products for the bus. POLITICAL: [Here I can offer much better experience having watched the computer business for some time.] If you've watched the computer business in recent years you've probably taken note of the patterns of success and failure for various lines and individual products. One of the most striking patterns in my mind is that of open architecture systems versus closed: DEC UNIBUS & QBUS, S100 hobby bus, Apple II, IBM PC, XT and AT, VME, etc. on the open side were almost instantaneous hits. The ledger isn't 100% clear however with DEC SBI and BI, Apple Macintosh, and a few other closed systems doing extremely well. But if memory serves me, the Apple Macintosh had an uphill battle before it finally achieved the popularity it now enjoys. I don't have enough background with the DEC SBI and BI to comment on any disadvantages they may have experienced because of closed designs - they may well have succeeded simply because people were already locked into the VAX architecture and needed more CPU ... In any case, there's a popular feeling among [some] manufacturers that if they design their systems around proprietary architectures, then they can get 100% of the resulting after market business and thereby achieve better profits. It's my contention that this is a false premise. I believe that if you design systems around open standards, you may not get 100% of the after market business, but because there will be so many third party options available for your systems, the overall sales of your system will more than make up for any percentage after market loss. This note is already too long, so I'll end it here. Casey
kjeld@iesd.dk (Kjeld Flarup) (04/24/89)
I only have a two demands for busses. 1) They must be able to address at least 4Gbyte 2) There must be a fibre optics net interface. It is my definite opinion that Disk, Printer and all other I/O must be standarized to a net protocol. Using fibre optics for this gives at least the following advantages. - No direct electrical connection between equipment, meaning fewer errors and cheaper installation. - A simple plug, gives possiblity any kind of IO. - Lesser radio noise escaping from computers. - High speed on a serial standard. Note that RS-232-C with Xon/Xoff is about the only communications standard, that really is a standard. I would also like to note the INMOS transputers, which currently is based on serial links between modules. Moving into parallel computing worlds displacing parallel electric busses with serial fibre optics seems nessecary. Kjeld Flarup Christensen