Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!uunet!crdgw1!sixhub!davidsen From: davidsen@sixhub.UUCP (Wm E. Davidsen Jr) Newsgroups: comp.sys.ibm.pc.hardware Subject: Re: Cyrix CX-803D87-20 Coprocessor Message-ID: <1981@sixhub.UUCP> Date: 1 Oct 90 00:14:50 GMT References: <1969@sixhub.UUCP> <1990Sep30.160127.3442@rodan.acs.syr.edu> Reply-To: davidsen@sixhub.UUCP (bill davidsen) Organization: *IX Public Access UNIX, Schenectady NY Lines: 103 In article <1990Sep30.160127.3442@rodan.acs.syr.edu> amichiel@rodan.acs.syr.edu (Allen J Michielsen) writes: | At risk of starting a flame fest (pump, pump, pump, light).... | When cyrix (or anyone) claims a 3X or whatever speedup with stuff og this type | they REALLY are saying (in this case) that the average (or some measure) | 80x87 instruction is performed in 1/3 the number of clock cycles. Right. I haven't seen a claim that high, but that's how they get their faster. I have seen notes that certain chosen instructions were 3X faster, but not even a claim that the overall performance was that much better. | as a user, you will see virtually NO difference (generally) once the entire | system is assembled. That depends on your use and applications. | A typical application such as lotus or autocad does | use the 80x87, but a. the instruction mix they use almost certainly doesn't | line up with the one the claim was made using, & b. the stream of 80x87 | instructions is intermixed. For 1 80x87 instruction, there may be 50 or | 100 80x86(8) & memory:I/O related instructions. Using this model, with | the original claim, and some wild ass extrapolations, it should be clear | that while the 80x87 is important, just popping one in (or changing it) | won't ever lead to ANYTHING even remotely approaching nX speed ups. I agree that those applications are not going to notably benefit. I disagree that these are typical. I would claim that there simply is *no* typical application today (if there ever was) and that basing any claim for or against the need for additional hardware is either foolish or misleading, depending on whether you believe it yourself. | As a matter of fact, with lotus and any speadsheet I have (distributed | or my own), the difference with or without a 80x87 is som small I can't | even measure it. On the other hand, in autocad, the difference between | with and without a 80x87 can be signifigant and easily measured. But this | number is much bigger (by a signifigant factor) than that even remotely | possible with a more effecient 80x87. I agree completely, for these applications you would see little improvement from a faster math chip. You would se some, particularly if you did a lot of transcendental functions, but not much. | Further, in a followup post to this original, one MAY get the idea | that more performance can be achieved by simply popping in a faster | chip speed in a slower socket. I hope you get that idea, that's what I said. | This is NOT the case, the actual system | clock speed is like a traffic cop. Everything goes that speed in that system. You were right a few paragraphs back. The better chips perform the operation in fewer cycles. Therefore they do more operations in the same time, or run the same program in less time. Without a change in clock speed. Your first statement was correct, this one is wrong. | So, when a 100 Mhz chip is put in a 4.77 Mhz system, it will give very very | very nearly exactly the exact same performance as the same model chip with | a 4.77 Mhz rating. (There can be a terribly small differences but they | are so small it is nearly impossible to measure.) In this case, the | reference being made is to a fundamental change in chip design. This | change was implemented at a higher speed rating, and is responsible for | a measureable performance increase if/when that chip is used in a slower | system. As I read it, you just corrected yourself. If you use a redesigned chip at the same speed it can run programs in fewer cycles (= less time). I have seen some real programs (ie not benchmarks) which ran 25 times faster with a coprocessor. I make no claim that these are typical, but while 100 sec is a lot better than 2500 sec (that's the ballpark for this program), knocking 30% off that is going to be quite useful. I will believe the 3X when and if I see it. I have seen figures taken by someone who traded a 387 for an IIT and got about that 30%. Examples of programs which might do this kind of thing are star charts, ray tracing, and calculation of complex orbits. There are people who do that kind of thing, even for fun. Also fractals if float is being used. | It also is my understanding that the redesign of the 80387 chip by | intel that was implemented at the 33 Mhz speed, has been reapplied | across the board to all 80387's now in porduction. Can anybody give | reference to either confirm or deny that rumour ? (Anybody listening | at intel ? -- I guess silence may be a answer also...) I have no more solid information than you. My pet rumor says the process and masks for the 25MHz has higher yield and therefore is still in use for the chips which are slower rated. Note I have no more confirmation than you, and if I get it it will probably be under non-disclosure and I couldn't say so anyway. There are applications which benefit from 860s and vector processing boards. Every user must be quite careful to understand what resources s/he uses before buying hardware to provide more of same. On that I agree with you completely and without reservation. I usually stop there trying to tell someone else what will and won't be effective. -- bill davidsen - davidsen@sixhub.uucp (uunet!crdgw1!sixhub!davidsen) sysop *IX BBS and Public Access UNIX moderator of comp.binaries.ibm.pc and 80386 mailing list "Stupidity, like virtue, is its own reward" -me