franka@mmintl.UUCP (04/15/87)
In article <6208@bu-cs.BU.EDU> bzs@bu-cs.UUCP (Barry Shein) writes:
... that with a hypothesized 50 to 1000 x speed improvement, we may have
more processing power than we know what to do with.
I think he's wrong. Consider what happens when you add natural language and
voice recognition to the user interface of every program. When you are
supporting a 10,000 by 10,000 pixel screen in graphics mode. When large
amounts of real-world knowledge are to be built in to a program to enable
intelligent behavior.
Now, these things require varying degrees of software development; but it
seems likely to me that each will also use up immense amounts of computing
power. No one has yet tried to develop such things for market, because the
hardware required to support them is not available yet. (Or if they *have*
tried, they have failed because ...)
This is relevant to the "who needs giga-bytes of memory?" controversy, too.
These kinds of applications will require huge data storage, too. This is a
general phenomenon; for a given level of functionality, there is a tradeoff
between speed and space, but enhanced functionality requires more speed *and*
more memory.
Frank Adams ihnp4!philabs!pwa-b!mmintl!franka
Ashton-Tate 52 Oakland Ave North E. Hartford, CT 06108
bzs@bu-cs.BU.EDU (Barry Shein) (04/17/87)
Posting-Front-End: GNU Emacs 18.41.4 of Mon Mar 23 1987 on bu-cs (berkeley-unix) No Frank, you misunderstand me. You say "sure we'll need gigaflops, we'll need it to do natural language and voice recognition..." That's my whole point. We don't know how to do all that and I don't believe (completely, trying not to be too dogmatic here) that it's the lack of cycles that's holding us back. I simply mean what I said. That other than the increasingly smaller percentage of number crunchers out there (smaller because the computer using population is growing [eg. PCs] and not in the cruncher area) we won't have applications for the majority of users which will utilize all those cycles (remember folks, I'm talking like 100MIPs desktops.) To put it another way, you have no proof (not getting stuffy here) that what's required to do natural language is more cycles. It seems plausible, but why aren't there any good software systems running on Crays (eg, or pick your favorite appropriate high-end box.) I'm just saying it's not obvious the lacking thing is cycles. Something to do with hardware innovation requiring linear effort while software innovation requiring exponential effort, but I babble, nothing to back that up. -Barry Shein, Boston University
kent@xanth.UUCP (04/18/87)
Barry, I think a bit more imagination is in order in this discussion. If the processing speed is provided cheap enough, it is easy to envision home uses for the Teraflop micro with the Terabyte memory chip. ;-) Examples which come to mind immediately: Flying your home computer in real time through the Mandelbrot set, recomputing at the 60 Hz frame rate, 1280 by 1024 by 24 bits display time, 2048 iterations deep. Robot wars with real time ray traced graphics for the kiddies, same resolution (you didn't think I was going to buy two of those things at a week's pay apiece, did you! ;-). Landscape design 101, home instrucion course, with fractal rocks, texture mapped fence boards, rule grown trees and shrubs, architectural database buildings, all ray traced in real time as the user moves stuff about in search of a passing grade. Stock investment with a quick AI rule based search of all the stocks on all the markets, based on daily highs and lows for each for the last five years, looking for a place to put the family nest egg until tomorrow night. A full inverted text database search of all the periodicals published in all areas of science this century, for a high school "civics of science" project. Real time scanning all 2000 available TV channels, looking for a show with redeeming social value, using another AI program with a built in optimistic streak. ;-) Home weather prediction using data from the new 1 km worldwide grid. And the list goes on. The point being that work expands to consume the resources available for it; always has, always will. ;-) Kent. -- The Contradictor Member HUP (Happily Unemployed Programmers) // Also // A Back at ODU to learn how to program better (after 25 years!) \\ // Happy \// Amigan! UUCP : kent@xanth.UUCP or ...{sun,cbosgd,harvard}!xanth!kent CSNET : kent@odu.csnet ARPA : kent@xanth.cs.odu.edu Voice : (804) 587-7760 USnail: P.O. Box 1559, Norfolk, Va 23501-1559 Copyright 1987 Kent Paul Dolan. How about if we keep the human All Rights Reserved. Author grants free race around long enough to see retransmission rights, recursively only. a bit more of the universe?
tpmsph@ecsvax.UUCP (Thomas P. Morris) (04/18/87)
In article <6654@bu-cs.BU.EDU>, bzs@bu-cs.BU.EDU (Barry Shein) writes: > To put it another way, you have no proof (not getting stuffy here) > that what's required to do natural language is more cycles. It seems > plausible, but why aren't there any good software systems running on > Crays (eg, or pick your favorite appropriate high-end box.) I'm just > saying it's not obvious the lacking thing is cycles. Barry, not being an expert on Natural Language Processing or Voice Recognition, I agree that it _may_ not be the lack of cycles (or memory, for that matter) which is holding back these technologies. However, it wouldn't seem very cost- effective to even do research on these topics using a Cray, or any other favorite high-end box (perhaps a Connection Machine? ;-)). To make the _software technology_ widely available, I'd think that a desk-top 100MIP, 4Gb machine (say, $10K, 5yrs from now :-> ) would make the research and the realization of these technologies more cost effective.
johnson@uiucdcsp.cs.uiuc.edu (04/18/87)
There are a number of programming styles that result in programs that are small, easy to understand, and slow. Logic programming and constraint programming both fit into this catagory, and functional programming probably does too, except that in the last few years, we have seen a number of efficient implementations of functional programming languages. My favorite language is Smalltalk, and its sudden popularity is due in no small part to the increase in speed of cheap processors. However, it is still too slow for many applications. Other interesting languages, like SETL, suffer from the same problem. What about 3D graphics? I've used an IRIS workstation, and I'm sure that you could sell millions of them if they were only a couple of thousand dollars apiece. 100 MIPS is probably about right for that kind of performance. We need fast processors just to search large databases. Naturally, you can keep lots of indexes, but that takes lots of work, and some things are hard to index. In general, writing efficient programs is hard work. It is easier to write clear programs and not worry about efficiency. We are still a long way from there. It won't be long until 100 MIPS will seem slow.
gerryg@laidbak.UUCP (04/19/87)
In article <6654@bu-cs.BU.EDU> bzs@bu-cs.UUCP writes: >I simply mean what I said. That other than the increasingly smaller >percentage of number crunchers out there (smaller because the computer >using population is growing [eg. PCs] and not in the cruncher area) we >won't have applications for the majority of users which will utilize >all those cycles (remember folks, I'm talking like 100MIPs desktops.) >To put it another way, you have no proof (not getting stuffy here) >that what's required to do natural language is more cycles. It seems >plausible, but why aren't there any good software systems running on >Crays (eg, or pick your favorite appropriate high-end box.) I'm just >saying it's not obvious the lacking thing is cycles. If you are saying that there are current machines which have enough cycles to do natural language processing, but we need some major breakthroughs in software; I agree completely. Of course by the time that happens, we will probably have pc's with as much or more power as present day super-computers. I suspect architectural advances will also play an important role; for example, connection machines, etc. I really like the idea of connection machines. In a few years, they will probably be commodity products like DRAM's are today. Which reminds me, I have a draft of Hillis' connection machine paper, and I am interested in learning about more current work. If anyone can give me pointers to papers, etc. I'd appreciate it. I am especially interested in how the communication chips work, and algorithms, programming, languages, etc. Thanks in advance. gerry gleason
jsp@b.gp.cs.cmu.edu (John Pieper) (04/19/87)
>This is relevant to the "who needs giga-bytes of memory?" controversy, too. >These kinds of applications will require huge data storage, too. This is a >general phenomenon; for a given level of functionality, there is a tradeoff >between speed and space, but enhanced functionality requires more speed *and* >more memory. Yes, and you are arguing for massive parallelism, for "active memories", in short, for something like the Connection Machine. A hyperactive RISC machine probably won't be any better at these problems than the CRAY. The complexity is too great. Besides, look at the most obvious problem: why pay megabucks for a big, fast, completely underutilized memory? Balancing computation and I/O is not enough; we must balance these with the amount of memory available.
bzs@bu-cs.UUCP (04/19/87)
Kent, I still claim almost everything in your list of things we -could- be doing if we only had extremely high performance PCs fall into one of the following categories: 1. Relatively trite (calculate the Mandelbrot set in my living room in real-time? ok, maybe someone out there is hot to do that, but I wouldn't try to build a market on it.) 2. Not obvious that hardware technology is the lacking component (all the AI examples you give, maybe they would run on a 10MIPs box just fine? How can you know? Why don't we run these applications today on expensive fast iron? Are the owners of said iron just not interested.) 3. Possibly more in the realm of other technologies, although in the right ball-park (eg. scanning huge data bases requires fast I/O as much if not more than CPU power, large numbers of cooperating CPUs might be more useful in such an endeavor than single massive CPUs, similar argument for image processing.) The problem, of course, is that we don't know how to use multiple CPU systems very well, so we go for the iron. My contention is simply that sheer CPU power has become vastly over-rated as a cure-all, and people are starting to realize this. Again, I am speaking for the vast majority of the users of computers, I fully accept there is a small percentage for whom no iron will be fast enough in the foreseeable future. I also do not think we have reached a plateau quite yet, but we are rapidly approaching some point where further increases (for most users) will be superfluous w/o significant software and other technolgy advances, maybe not even then. -Barry Shein, Boston University
webber@klinzhai.UUCP (04/19/87)
In article <6741@bu-cs.BU.EDU>, bzs@bu-cs.BU.EDU (Barry Shein) writes: > 3. Possibly more in the realm of other technologies, although > in the right ball-park (eg. scanning huge data bases requires > fast I/O as much if not more than CPU power, large numbers of > cooperating CPUs might be more useful in such an endeavor > than single massive CPUs, similar argument for image processing.) > The problem, of course, is that we don't know how to use multiple > CPU systems very well, so we go for the iron. There is plenty of parallelism in a single cpu, `wires' wait for no one :-) . Off board is always slower than on board -- off chip slower than on chip. If cpu's were fast enough, one would probably want to start using smarter management of the interchip communication channels. > My contention is simply that sheer CPU power has become vastly > over-rated as a cure-all, and people are starting to realize this. > > Again, I am speaking for the vast majority of the users of computers, > I fully accept there is a small percentage for whom no iron will be > fast enough in the foreseeable future. I also do not think we have > reached a plateau quite yet, but we are rapidly approaching some point > where further increases (for most users) will be superfluous w/o > significant software and other technolgy advances, maybe not even then. Well, computers in general are irrelevant to most of humanity. However, fast cpu's can offer things of interest to the non-computer specialist. For example, current computer chess games are really bad in the endgame -- there is a fair amount of indication that fast hardware helps more than clever programming (although maybe no one has been clever enough yet :-). Fast cpu's could generate interactive universes -- roaming around in high-dimensional surface graphs would give much insight into the models of scientists and engineers. If cpu's were fast enough, everything you sent to disk could go encrypted -- adding significantly to the overall security of computer users (of course, fast cpu's will weaken the security of slow cpu users -- which sounds like a classic example of a market bootstrapping itself). Faster cpu's should mean greater reliability as things like array-bounds and pointer types can get checked at execution time. Fast cpu's can run algorithms to squeeze the maximum bandwidth out of slow i/o channels. Of course, people who like to manipulate computers directly will win even bigger. How would you like your compiler to be able to run global error correction algorithms instead of local algorithms that go crazy after the first 4 errors? How would you like to be able to use arbitrary context-free grammars to parse your favourite personal language rather than have to squeeze everything into LALR(1)? How would you like to have the computer manipulate numbers as flexibly as you do by hand (64bit arithmetic? -- give me bignums -- roundoff errors are a pain)? How about cpu's fast enough to allow your screen editor to compute the optimal update to your terminal screen? or fast enough to globably anti-alias output on your graphics device? How would you like your computer to be fast enough that your could run extensive code optimization algorithms that generate code fast enough so that on your machine it runs fast enough to allow you to run extensive code optimization algorithms ... Well there is much more, but I think you get the idea. --------------- BOB (webber@aramis.rutgers.edu ; BACKBONE!topaz!webber)
jtr485@umich.UUCP (Johnathan Tainter) (04/19/87)
In article <6654@bu-cs.BU.EDU>, bzs@bu-cs.UUCP writes: > > No Frank, you misunderstand me. You say "sure we'll need gigaflops, we'll > need it to do natural language and voice recognition..." > > That's my whole point. We don't know how to do all that and I don't > believe (completely, trying not to be too dogmatic here) that it's the > lack of cycles that's holding us back. No where in his posting did he say that limited cycles was the ONLY thing holding back these applications. > I simply mean what I said. That other than the increasingly smaller > percentage of number crunchers out there (smaller because the computer > using population is growing [eg. PCs] and not in the cruncher area) we > won't have applications for the majority of users which will utilize > all those cycles (remember folks, I'm talking like 100MIPs desktops.) Even if these applications don't need 100MIPs for themselves, when you have a voice synthesizer, a voice input system, a background music generator, a video digitizer input system, a posture monitoring video digitizer system, an evolving backdrop for your desktop metaphor and your application running at once you are going to eat mucho processor time. But, you might say, "Noone would do that!" And I say, noone will do that with what we can offer today (i.e. put in lots of special hardware). But give them 100MIPS to work with and find out just how much people will do. Also find out how much people will demand once they know it can be done. There are already people clammering for voice input systems, and those are still in the toy phase. There is one more field where mucho processor in a small package is going to be required. Robots. Classic SF style robots, not the misnamed things they use in factories. Also prostheses. Undoubtedly you have heard the axiom "Any program will expand to fill all available space.". If you take this for the obvious analogy to gases then we have a long way to go before the distribution of programs throughout the available processing space is sparse. And until we get it sparse we will never get computers to be another unremarked tool like a doorknob. Until we eliminate the need for programmers and software designers/engineers (like me) we have not taken the computer far enough. No man is doing his job properly unless he is working to make himself obsolete. > -Barry Shein, Boston University --j.a.tainter
bzs@bu-cs.BU.EDU (Barry Shein) (04/23/87)
Posting-Front-End: GNU Emacs 18.41.4 of Mon Mar 23 1987 on bu-cs (berkeley-unix) From: jtr485@umich.UUCP (Johnathan Tainter) First you quote me...(in re natural language processing) >> That's my whole point. We don't know how to do all that and I don't >> believe (completely, trying not to be too dogmatic here) that it's the >> lack of cycles that's holding us back. Then you say... >There is one more field where mucho processor in a small package is going to >be required. Robots. Classic SF style robots, not the misnamed things they >use in factories. Also prostheses. Sigh...you really are just going to ignore everything I am saying and go scramble for another identical example. Do you know of any SF style robots which are currently not able to be put into production due to a lack of cycles? Any prostheses? Any that work slowly or off of a ridiculously (physically) huge machine for the application (eg. a prosthesis being driven by a Cray-2)? No. So this is exactly the same point I answered in your previous quote of my text about natural language processing. You don't *know* that SF style robots or prostheses will take enormous amounts of MIPs, you just believe it to be true for some reason without having ever seen a single example. Maybe you're right, we'll just have to wait and see (how long?) >Undoubtedly you have heard the axiom >"Any program will expand to fill all available space." Yes, but that is not an axiom, it is a shibboleth. An axiom is something you can prove. You can't even demonstrate one common example given the next 5 years of processor development. Yes, we have seen it in the past and even in the present. But it's not obvious that this extrapolates into the future in a situation like this. My favorite example is we saw automobiles in the beginning of the century go from (max) 10MPH to 30MPH to 100MPH. But after that first (ca.) 30 years of development extrapolating the curve became ridiculous. The cost of building a car and designing a road system which would transport the average person at 300MPH became ridiculously out of reach. Sure, there's a flaw here in that lots of MIPs won't endanger your life (maybe, I could argue that one also, but it's more socio-technical) but the point where you cross over beyond cost/benefit might exist in a similar fashion. And maybe some day we will build 300MPH passenger cars, *maybe*. You'll get into more and more expensive technology to build this 1000MIP desktop and there just won't be enough of a market for it. Your potential customers will say "gee, I dunno, I can't keep my 100MIP PC you sold me last year busy, you have any applications to justify the cost of this new box?" And the answer may be "no". (don't quibble the numbers, insert 50MIPs and 100MIPs in there.) All I am saying is, not that we will never, or that no one will ever need huge MIPs. I am simply saying that we will probably run out of useful applications to justify them IN MOST CASES. If it takes 1000MIPs to run an SF robot we probably will have had 1000MIPs for 20 years before anyone demonstrates the need. Will you still be in business? It's really not all that subtle, is it? Or am I just really riling the more parochial out there? Perhaps the problem is that people tend to confuse quantity with quality, anyone remember the "cup and a half of flavor" ads? -Barry Shein Pae. T: <naturbubu
pase@ogcvax.UUCP (04/23/87)
In article <bu-cs.6872> bzs@bu-cs.BU.EDU (Barry Shein) writes: > [...] >Yes, but that is not an axiom, it is a shibboleth. An axiom is >something you can prove. I beg your pardon? According to Webster (1942 ed.) ax'iom (ak'si.um), n. [From L., fr. Gr. *axioma*, fr. *axioun* to think worthy, fr. *axios* worthy.] 1. An accepted maxim. 2. *Logic & Math.* A statement of a self-evident truth; thus, the statement that the whole is greater than any of its parts is an *axiom*. 3. An extablished principle which is universally received; as, the *axioms* of science. -- Doug Pase -- ...ucbvax!tektronix!ogcvax!pase or pase@Oregon-Grad (CSNet)
ram@wb1.cs.cmu.edu.UUCP (04/23/87)
As to "what are we going to do with all those cycles?", I find the "10 mips is enough" arguments to be pretty unconvincing. The "cycles don't limit X, we just don't know how to do X at all" argument is somewhat flawed in that it ignores the possibility that there may be fairly easy brute-force solutions that have been ignored due to computational infeasibility. It is true that such an algorithm, once developed, could run slowly on generally avilable hardware, or less slowly on today's supercomputers. This argument ignores the problem of getting the software developed in the first place. Where are all the researchers who have unlimited supercomputer access to develop these algorithms in real time? Consider that the algorithms might be initially developed in a form that is several orders of magnitude slower than eventual usable versions. This means that your researcher with his personal supercomputer would still find development painfully slow. Hans Moravec (who works on mobile robots at CMU) gave a talk in which he argued fairly convincingly that massive cycles make lots of robotic problems simpler. He had an interesting slide which plotted the log of data processing capacity on the Y axis and the log of data storage capacity on the X axis. He then plotted the capicities of various animals and computers. Interestingly they all fell on about the same line, meaning that the ratio of storage capicity to processing capacity is roughly constant and is similar between computers and animals. The bad news for A.I. and robotics comes in when when you notice that a ~1mip computer is about equal to a housefly... I believe that the storage and processing capacity of a human brain was estimated to be three orders of magnitude greater. He also plotted in a linear time scale along the bottom. He included data points for electromechanical and mechanical data processing and calculating machines, and showed that the exponential trend predated electronic computers, going back before the turn of the century. In order to deomstrate a way in which having lots of cycles could make life easier, he discussed ways of representing a robot's "mental map" in some detail. These are data structures used to answer the question "Is there anything there?" and "How can I get from here to there without running into anything?". Initial attempts used geometric models, representing objects as polygons and paths as lines. These methods weren't too successful because they can't deal with uncertainlty very well; the results of the robot's sensors tend to have a great deal of uncertainty. A more successful but computationally intensive method involves breaking space up into a grid and representing the certainly of an object being in each location. This representation is very good at dealing with sources of uncertainty since you can combine the results of various inputs at each location and can also apply transformations to smear out the probabilities to represent things such as uncertainty in how much you have moved. A successful path plotting algorithm is based on this representation. It basically looks for the "water running downhill" path, if the certainty of an obstruction is considered to be height. This algorithm also seems to be "embarrassingly parallel", which is a good thing since we probably will running into diminishing returns on uniprocessors pretty soon. Comparisons with automobiles (the classical mature technology) are spurious. In fact it seems that automobiles are a qualitatively different kind of technology. In the entire course of aotomobile development we have only seen one or two orders of magnitude improvement. I can think of a few essential differences: -- Physical limits. The 60,000 mph automobile simply isn't possible. If you made one, it wouldn't be an automobile at all, it would be a spacecraft, which brings us to: -- Definition. The concept of an automobile is much less flexible than that of a computer. An automobile is defined by its design (ground vehicle with ~4 wheels that carries a few passengers) rather than what it does (human transportation). In contrast, almost any information processing device is called a computer. Some of the parallel architectures being considered seem more different from an IBM PC than a 747 is from a VW. [500 people at 700 mph is 350,000 person-miles per hour, as opposed to perhaps 240 for a VW.] -- External limitations, in particular scale. An automobile is required to carry people, and they are rather inflexible. If transportaion designers could minaturize people and their environment, redesigning them to run in liquid helium, then much more impressive improvements in transportation might be attainable. The use of "speed" to refer to information processing capacity also introduces a compelling but largely spurious analogy with automobiles. A faster automobile is harder to use and only gets you there faster. In contrast, a faster computer probably won't respond to interactions any faster; instead it will be easier to use and will let you do things that simply weren't feasible before. Rob
bjorn@alberta.UUCP (04/24/87)
It seems to me like carrying water into an already flooded cellar to post this reply, but the sentence about axioms is what really brought the water level to the ground floor. In article <6872@bu-cs.BU.EDU>, bzs@bu-cs.BU.EDU (Barry Shein) writes: >>Undoubtedly you have heard the axiom >>"Any program will expand to fill all available space." > >Yes, but that is not an axiom, it is a shibboleth. An axiom is >something you can prove. Please, please!?! We've seen enough perversion of perfectly reasonable CS concepts by the PC people in years gone by (not to mention re-inventing each and every shape of wheel in existence), let's not start on mathematics (or general use English for that matter) already. An axiom is a given even though mathematicians do not select them frivolously and sometimes some pretty astounding results have been proven equivalent to specific axioms for particular theories (Axiom of choice == Zorn's lemma! [there are more results that are "==" to these] anyone? [wrong group!!]). > My favorite example is we saw automobiles in the beginning of the > century go from (max) 10MPH to 30MPH to 100MPH. But after that first > (ca.) 30 years of development extrapolating the curve became > ridiculous. This is a ridiculous example, as you note later on (wonder why it's your favorite). The whole point is that there are some fairly well entrenched physical laws (barriers) to contend with. This analogy, favored as it may be, is therefore less than useless. It's not even true, for indeed we have gone beyond x mph autos long ago with those devices that are commonly referred to as airliners, bullet trains, etc.. > The cost of building a car and designing a road system > which would transport the average person at 300MPH became ridiculously > out of reach. See above. > Sure, there's a flaw here in that lots of MIPs won't > endanger your life (maybe, I could argue that one also, but it's more > socio-technical) but the point where you cross over beyond > cost/benefit might exist in a similar fashion. And maybe some day > we will build 300MPH passenger cars, *maybe*. The crucial term in the above is "cost/benefit", I don't think we have any argument with a one man non-existent market, although I must say that taking a flight through a Mandelbrot set in real time is not my fantasy for an ideal Saturday night. > You'll get into more and more expensive technology to build this > 1000MIP desktop and there just won't be enough of a market for it. What we have continually (so far) is a see-saw (sp?) of processor and memory technology. Wasn't far back that you didn't need any tricky or expensive designs to get away with 0 wait state memories in typical state of the art microprocessor systems, wasn't long before that when you had to have cache, etc. to keep your processor busy. Look's like were stuck with some pretty expensive memory system designs for the next little while, if we want to squeeze out every last ounce of performance from a processor that is. > Your potential customers will say "gee, I dunno, I can't keep my > 100MIP PC you sold me last year busy, you have any applications to > justify the cost of this new box?" And the answer may be "no". (don't > quibble the numbers, insert 50MIPs and 100MIPs in there.) Have you seen how OS's are being implemented these last few days. Did you ever hear about dividing the iq of dumbest number of a committee by the number of people on said committee. If recent experience is any indication OS's are going to take a good slice of whatever cycles are present. Prime opportunity for slick OS fellows I'd say. > All I am saying is, not that we will never, or that no one will ever > need huge MIPs. I am simply saying that we will probably run out of > useful applications to justify them IN MOST CASES. If it takes > 1000MIPs to run an SF robot we probably will have had 1000MIPs for 20 > years before anyone demonstrates the need. Will you still be in business? Yes and I'm just nagging when I bitch (silently and privately) about the way some people waste the cycles in their Sun's drawing frivolous pictures 24 hours a day such that I can only get 80-90 % cpu out of same Sun's instead of the 10000 % I really want. > It's really not all that subtle, is it? Or am I just really riling the > more parochial out there? No it's not subtle at all. Give me and thousands of others all the cycles you have left over (preferably all at once mind you) and we'll be as flushed as cooked lobsters from our good fortune. I have plenty of applications that can use absolutely any amount of cycles that you give me, and no they're not necessarily graphics, NLP, seismic or weather [or other PDE stuff] related. > Perhaps the problem is that people tend to confuse quantity with > quality, anyone remember the "cup and a half of flavor" ads? No, was that on TV by any chance? No wonder I missed it. Bjorn R. Bjornsson {ubc-vision,ihnp4,mnetor}!alberta!bjorn
bzs@bu-cs.BU.EDU (Barry Shein) (04/25/87)
Posting-Front-End: GNU Emacs 18.41.4 of Mon Mar 23 1987 on bu-cs (berkeley-unix) >In article <6872@bu-cs.BU.EDU>, bzs@bu-cs.BU.EDU (Barry Shein) writes: >>>Undoubtedly you have heard the axiom >>>"Any program will expand to fill all available space." >> >>Yes, but that is not an axiom, it is a shibboleth. An axiom is >>something you can prove. > >Please, please!?! We've seen enough perversion of perfectly >reasonable CS concepts by the PC people in years gone by (not >to mention re-inventing each and every shape of wheel in >existence), let's not start on mathematics (or general use >English for that matter) already. OK, ok, I screwed that up. An axiom is something you take as true, a given, tautological or otherwise self-evident. Does this mean everyone complaining about that agrees that "any program will expand to fill all available space" is an axiom? Or is it a shibboleth as I claimed? Or is everyone just jerking off in public? Maybe conversations on these lists would be more productive if people would try to address the issues with substantive debate instead of trying to point out minor slip-ups of no real consequence to the overall argument. I think my point was clear enough, that statement is not an axiom in any sense of the word (my misuse or your correction), it's a homily, a shibboleth, actually a humorous joke called "Parkinson's Law" that's analogous to Murphy's Law and many other amusing self-flagellations that engineers use to try to laugh at their errors. It has no place in a serious argument, it sheds no light on anything. -Barry Shein, Boston University
franka@mntgfx.MENTOR.COM (Frank A. Adrian) (04/28/87)
In article <6872@bu-cs.BU.EDU> bzs@bu-cs.BU.EDU (Barry Shein) writes: >My favorite example is we saw automobiles in the beginning of the >century go from (max) 10MPH to 30MPH to 100MPH. But after that first >(ca.) 30 years of development extrapolating the curve became >ridiculous. The cost of building a car and designing a road system >which would transport the average person at 300MPH became ridiculously >out of reach. Sure, there's a flaw here in that lots of MIPs won't >endanger your life (maybe, I could argue that one also, but it's more >socio-technical) but the point where you cross over beyond >cost/benefit might exist in a similar fashion. And maybe some day >we will build 300MPH passenger cars, *maybe*. > > -Barry Shein, Boston University I agree completely with the above argument. Lets talk about constraints as it applies to I/O bandwidth. Let's postulate a machine with 4 Tbytes of memory and a GAF (Gawd Awful Fast) CPU with unlimited bandwidth. Lets look at the new bottleneck in the system. Here I am with a (small :-) 2Tb object (say something dumb like an atmospheric model (which someone said he wanted to do at home with data fed in from a worldwide sampling network)), and I want to send it to a friend. Lets look at some shirtsleeve calculations. Let's see how long I have to wait to send this data at different baud rates: baud time ---- ---- 9600 17Gs (this is GIGA, folks) 50K 320Ms (getting better, but still have to wait a while) 50M 320Ks (down to about 9 hours) 50G 320s. (right ballpark) OK, now we're talking about the full effective bandwidth of a good size communications satellite (using several parallel channels), and it still takes ~5 minutes to dump half your physical memory. Now I don't think that we have enough bandwidth to support all of the I/O bandwidth we are going to need, and that's the communication infrastructure which needs to be in place before we even get enough data to need these sizes of memories and CPU speeds. Also, if 4 Tbytes is a physical memory, what's a storage device going to look like? Electron beams fired just outside the event horizon of quantum black holes to form storage loops? Gives a whole new meaning to the phrase "spin-up". Just food for thought... Frank Adrian Mentor Graphics, Inc.
franka@mmintl.UUCP (Frank Adams) (04/29/87)
In article <6654@bu-cs.BU.EDU> bzs@bu-cs.BU.EDU (Barry Shein) writes: >No Frank, you misunderstand me. You say "sure we'll need gigaflops, we'll >need it to do natural language and voice recognition..." > >That's my whole point. We don't know how to do all that and I don't >believe (completely, trying not to be too dogmatic here) that it's the >lack of cycles that's holding us back. >... >To put it another way, you have no proof (not getting stuffy here) >that what's required to do natural language is more cycles. It seems >plausible, but why aren't there any good software systems running on >Crays (eg, or pick your favorite appropriate high-end box.) I'm just >saying it's not obvious the lacking thing is cycles. We can leave the Crays alone; those machines' time is too valuable to waste on making their user interface more friendly, especially since the kind of applications they are used for are not interface intensive anyhow. There are natural language database systems on mainframes. My impression is that they suffer about equally from lack of computing power and lack of software capability. They aren't half bad, though. (By the way, this opinion is based on seeing canned demos, but no real hands on experience, so take it for what it's worth.) It is also worth noting that in many cases, *either* more sophisticated software or more powerful machines will get you what you want. For example, I think it is clear that there are adequate pattern recognition programs (voice or graphics) which don't run fast enough on hardware with a price low enough to be feasible. We might be able to do better with better software; but we can certainly use what we have now if the hardware is two orders of magnitude faster for the same price. Frank Adams ihnp4!philabs!pwa-b!mmintl!franka Ashton-Tate 52 Oakland Ave North E. Hartford, CT 06108
pase@ogcvax.UUCP (04/30/87)
In article <franka.619> franka@mntgfx.UUCP (Frank A. Adrian) writes:
-[...]
-Let's see how long I have to wait to send this data at different baud rates:
-
-baud time
----- ----
-9600 17Gs (this is GIGA, folks)
-50K 320Ms (getting better, but still have to wait a while)
-50M 320Ks (down to about 9 hours)
-50G 320s. (right ballpark)
-
- [...]
-
-Frank Adrian
-Mentor Graphics, Inc.
Don't forget the box of optical disks sent via some neighbor kid on a bicycle
(or Federal Express, etc.).
--
Doug Pase -- ...ucbvax!tektronix!ogcvax!pase or pase@Oregon-Grad (CSNet)
eugene@pioneer.arpa (Eugene Miya N.) (05/04/87)
In article <2119@mmintl.UUCP> Frank Adams writes: >We can leave the Crays alone; those machines' time is too valuable to waste >on making their user interface more friendly, especially since the kind of >applications they are used for are not interface intensive anyhow. WRONG....;-)! >There are natural language database systems on mainframes. > >Frank Adams ihnp4!philabs!pwa-b!mmintl!franka >Ashton-Tate 52 Oakland Ave North E. Hartford, CT 06108 Frank, and every one else: While I don't like the word friendly, the end user's time is typically more valuable that the machine, even $20M machines. The end goal is the comprehension of the science, and if you can't get your results back in time, why use a computer? Consider the days of the 24-hours weather forecast which took 27-hours to run on a 7600. Well, I guess it makes it easy to check your results. ;-) Make them more friendly. I don't think we really have good NL systems. Geez, this new group has really grown. I wonder why. I don't want to unscribe. Lots of the stuff is useful to listen to. From the Rock of Ages Home for Retired Hackers: --eugene miya NASA Ames Research Center eugene@ames-aurora.ARPA "You trust the `reply' command with all those different mailers out there?" "Send mail, avoid follow-ups. If enough, I'll summarize." {hplabs,hao,ihnp4,decwrl,allegra,tektronix,menlo70}!ames!aurora!eugene
howard@cpocd2.UUCP (Howard A. Landman) (05/08/87)
In article <franka.619> franka@mntgfx.UUCP (Frank A. Adrian) writes: >Let's see how long I have to wait to send this data at different baud rates: >baud time >---- ---- >9600 17Gs (this is GIGA, folks) >50K 320Ms (getting better, but still have to wait a while) >50M 320Ks (down to about 9 hours) >50G 320s. (right ballpark) In article <1263@ogcvax.UUCP> pase@ogcvax.UUCP (Douglas M. Pase) writes: >Don't forget the box of optical disks sent via some neighbor kid on a bicycle >(or Federal Express, etc.). Yes, even today, for large amounts of data, the speed and cost of moving a physical medium can be much better than that obtained by transmitting pure information. It's instructive to compute the baud rate of a mundane 1600 BPI tape flown from NY to SF in 6 hours. And of course, you can easily increase the effective rate by an order of magnitude by simply sending 10 tapes at once. For Doug's case, assume an optical disk means a CD, holding 600 MB. For micro users, the cost of downloading a "free" program from, say, Compuserve may exceed the cost of buying a disk with that program on it! -- Howard A. Landman ...!intelca!mipos3!cpocd2!howard howard%cpocd2%sc.intel.com@RELAY.CS.NET (it worked for RMS!) "My copyright left, but my copyleft was right!"
adam@misoft.UUCP (05/13/87)
In article <6872@bu-cs.BU.EDU> bzs@bu-cs.BU.EDU (Barry Shein) writes: >>Undoubtedly you have heard the axiom >>"Any program will expand to fill all available space." >Yes, but that is not an axiom, it is a shibboleth. An axiom is >something you can prove. No, an axiom is something that you hold to be true, without necessarily having to prove it. Typically people agree on a set of axioms from which to work, e.g. opposite angles being equal in Euclidean geometry. -Adam. /* If at first it don't compile, kludge, kludge again.*/
franka@mntgfx.UUCP (05/20/87)
>In article <franka.619> franka@mntgfx.UUCP (Frank A. Adrian) writes: >>Let's see how long I have to wait to send this data at different baud rates: >>baud time >>---- ---- >>9600 17Gs (this is GIGA, folks) >>50K 320Ms (getting better, but still have to wait a while) >>50M 320Ks (down to about 9 hours) >>50G 320s. (right ballpark) > >In article <1263@ogcvax.UUCP> pase@ogcvax.UUCP (Douglas M. Pase) writes: >>Don't forget the box of optical disks sent via some neighbor kid on a bicycle >>(or Federal Express, etc.). > In article <673@cpocd2.UUCP> howard@cpocd2.UUCP (Howard A. Landman) writes: >Yes, even today, for large amounts of data, the speed and cost of moving a >physical medium can be much better than that obtained by transmitting pure >information. It's instructive to compute the baud rate of a mundane 1600 BPI >tape flown from NY to SF in 6 hours. And of course, you can easily increase >the effective rate by an order of magnitude by simply sending 10 tapes at once. >For Doug's case, assume an optical disk means a CD, holding 600 MB. > >For micro users, the cost of downloading a "free" program from, say, >Compuserve may exceed the cost of buying a disk with that program on it! In all of this discussion I think my original point got lost (or maybe it was lost between my brain and the keyboard), which was that we have all of these companies wanking on about how many MIPs their processor has when in most cases, their I/O systems were the bottleneck in their last generation of machines. Actually, I don't see very much need for an even higher MIPs rate unless we can get he info in and out of memory in a timely fashion (and you'd have to have a preety small kid on a bicycle to cart the data out of memory). I know that it is tempting to keep adding more MIPs becuase it is a simple number for the average dweeb in the street to think he understands when the ol' marketing department comes around wanking (even though it is meaningless), but those of us who work with systems already know how little bang for the buck adding more MIPs seems to be getting these days. The question I'm throwing out to all of the great system architects out there is "What can be done to solve the problem of the apalling lack of I/O bandwidth found on every micro based uniprocessor system out there today?". Just wondering... Frank "Mr. Sensitivity" Adrian Mentor Graphics, Inc. These views in no way, shape, or form reflect those of my employer.