Path: utzoo!attcan!uunet!steinmetz!davidsen From: davidsen@steinmetz.ge.com (William E. Davidsen Jr) Newsgroups: comp.arch Subject: Re: memory speed & futurology Message-ID: <11978@steinmetz.ge.com> Date: 25 Aug 88 17:01:16 GMT References: <2179@ditmela.oz> Reply-To: davidsen@crdos1.UUCP (bill davidsen) Organization: General Electric CRD, Schenectady, NY Lines: 103 In article <2179@ditmela.oz> george@ditmela.oz (George michaelson) writes: | | The August edition of 'electronics' is about memory technology [ numbers followed by questions ] | (3) do the new speeds still look good alongside predicted clock speeds | for CPU or do we have another development lag here? | | there is another table in the mag showing possible access times | for existing 32bit cpus & available speeds from different memory | technology but it's hard to reproduce. It implies memory access | delay is one of many bottlenecks, I think these speedups might cure | it but only for existing clockspeeds. I think you've hit right on it. Fortunately, I believe that CPU speeds are not going up the way they have in the past five years (see below). | | (5) do we start to get 32/64Mb by default in our workstations? | does the opsys change its memory usage when that much memory | is around? I think you're a lot closer than that now. When we ordered some Sun 386i workstations with 4MB, we were told the delivery would be five weeks slower than the 8 or 16MB models. It seems that there is little demand for 4MB, and they only set the line for them every six weeks or so. Most of the stations we order have 8-16MB, and 4 is confining, even for single user. | | (6) with optical disk getting down to current access times for `real' | disk do they become standard or are the disadvantages still too | great? does unix sprout file version numbers? -I'm assuming WORM speeds | are similar to pre-recorded speeds here... It would seem that the access time on hard or optical disks is limited by rotational speed in the long run. You can add sectors and tracks to reduce the track to track seeks, but there is that limiting factor. Depending who you believe, it is obvious that the WORM has a higher limit than the magnetic disk, or exactly the opposite. There are good reasons for WORM drives, mainly legal. Things written on a WORM can't be changed, and therefore are more likely to be admissible as evidence. This exists in some courts now, but it's a topic for the legal group, not here. | | (7) dual ported memory & video rams: do they benifit too? I would expect them to benefit more, at least comparedto what we use now. Add ROMs to that list! With faster, cheaper, ROMs there will be more applications available for them. I would hope that EAPROMs would be here, too. ================ My view on where computers are going ================ There are some reasons why running CPUs at higher and higher speeds is not the most cost effective way to improve performance. As the speeds get higher the traces on the boards start to look at transmission line, you get radiation, standing waves, etc. These problems are soluble, but looking at the troubles today in meeting class B certification by the FCC, at some point it will be more cost effective to use other approaches. There are two well known ways to use slower speeds effectively, and both are in use now. The easiest is lowering the number of clock cycles per instruction. This is a benefit of RISC. There are some developments ongoing to have a mixed hard logic and microcode system. By making more of the common instructions hard logic, the code runs faster without trying to get all of the special purpose instruction in hard logic. The other is using a wider bus. This allows slower memory (and their busses) and takes advantage of larger register sets inside the CPU. Cache memory also reduces the effects of slow memory, and by placing cache on the CPU chip itself, it may be possible to run the CPU at higher speeds with fewer problems that if the faster signals were run on the backplane. It can be proven that there is a limit to how fast a computer may be, independent of the techniques used. There was an article by a physicist several years ago on this, and he quoted the limit. The problem is the speed of light. To reduce delays induced by the SOL requires making devices smaller. When wires becode extremely small they become statistical problems rather than conductors. If a wire is a few molecules in diameter (he quoted the values), putting an electron in one end does not insure that an electron comes out the other. Add to this having to use lower voltage to keep power down so the whole thing won't melt, and you hit a firm lower limit on size, and thereby performance. This does not preclude parallel processing for problems which have the right characteristics. The good news is that we are currently about 21 orders of magnitude from the limit. I can't guess what reducing the size of a CPU/memory system by that level would do for performance, but if you put one on my desk I'll report the benchmarks. Anyone who can find a copy of the original article, please supply more details. -- bill davidsen (wedu@ge-crd.arpa) {uunet | philabs | seismo}!steinmetz!crdos1!davidsen "Stupidity, like virtue, is its own reward" -me