owen@sbcs.sunysb.edu (Owen Kaser) (08/10/90)
Thanks to all those who responded to my posting. I received a few email messages in addition. Their contents were quite interesting, and I'm posting selected extracts. Names of contributors have been omitted, in case anonymity is desired. -- The original question -- Do timing specs for microprocessors and their support chips bear any relation to reality, and do industrial designers (need to) pay any heed to them? ---------------------------------------------------------------------- Remember LISP Machine, Inc.? Well the primary reason the company failed was because the LAMBDA processor was shipped to customers about 2 years before it was debugged. One of the many problems was running the chips over spec. To compensate for this, there were a few tapped delay lines that were adjusted at the factory until the machine ran. Needless to say, the machine was highly temperature sensitive. .... The people who said that we really didn't need to run within spec were fired real quick. ... If the manufacturers could guarantee that all their chips exceed the specs, they would change the specs. If you exceed the specs, you can no longer estimate the failure rate accurately. I certainly believe that exceeding the specs is poor engineering practise. ---------------------------------------------------------------- The published timing specs are what the manufacturer *guarantees* as a minnimum - that is, what is safe enough to get sued about if incorrect. Such data naturally includes a large margin of safety from what a chip can really do. ----------------------------------------------------------------- Statistically speaking, you were very very lucky. Even though we know our components always somewhat exceed our specs, we can't/won't tell you by how much and under what conditions. You will never achieve quality levels measureable in PPM unless you do a worst-case design. Certainly there are many engineers who take easier routes, but we consider it to be naive and irresponsible. We NEVER hear about problems related to our specifications from companies that design their systems correctly (e.g. HP, Apple, NCR, ...) but we often hear from those who don't. Something like: "we've been building this for years and we've never had a problem before" or "we where counting on some margin here". Sorry, but these guys don't get a lot of sympathy. -------------------------------------------------------------------------- Yep, its a good idea to pay attention to them. Depends on the manufacturer. If you get a part from Intel, their specs are SO BAD that you can bet the actual parts are gonna be MUCH better. If you buy a micro from NSC(Why would you do a silly thing like that..) you'd find that the specs tell an accurate story and if you violate them your gonna pay. ... Good engineers will design without violating timing/loading. This will allow you to margin the design over power AND temp at the same time. You SHOULD margin the designs before you let it out of the lab just to find the last of the flakies. You'd find that the main-frames are typically the best designs around, i.e. they will be the most conservative. They usually aren't done under the same constraints as designs at start ups, and they have more control over the parts supply quality. I've worked at a mainframe manufacturer as a design engineer....they have "rules" books that tell the designer which chips they can use, and what the chips performance is. These numbers are usually down rated from manufacturer specs(i.e. slower) so they will have an in-built margin. Main-frame companies also employ "circuits/package engineers" that just worry about parts quality, and design quality. All the timing violations your boss was so fond of ignoring would be disallowed as a valid design at a mainframe company(or any major company ;-) --------------------------------------------------------------------- Look at it this way: Is a manufacturer really going to tell you you can only run at x MHz when you can really run at x+1 MHz? It's not in their interests to do so; they want to be able to tell you they've got the fastest parts around. If they're good, they've done their homework and figured out the worst-case situations for everything. You may never hit that in a particular design, for various reasons, but unless you know what their criteria are it's not a good idea to go outside spec. -------------------------------------------------------------------------- ... I always try to live within the manufacturers worst case specs. About the only time I ever fudge on this is when I assume all gates in a single package will exhibit about the same characteristics, i.e. either they are all close to the best case or all close to the worst cast. This seems reasonable since they are all of the same grade silicon and all at the same temperature. Occasionally I run into problems where the worst case spec of one chip overlaps with the best case (i.e. minimum delay) of some other chip, such as when timing signals are propagating through parallel paths. That is when the above assumption comes in handy. ... I've found that most chips outperform their timing spec, but only at or near room tempera- ture. When the box heats up, watch out! ... <discussion of PC clones>... It never ceases to amaze me how flakey alot of these motherboards can be. It seems that a manufacturere will often take a board designed for 16 MHz and just plug in a 20 or 25MHz crystal. In fact, I've seen 25 MHz clock rates with only a 20MHz 80386, and I was told that was common practice. However, buying DRAM for these boards is an art. One vendor's 80 nsec. DRAM will work while another vendor's will not. They are obviously so close to the edge that the slight variation from vendor to vendor upsets the timing. ----------------------------------------------------------------
lewine@dg-rtp.dg.com (Donald Lewine) (08/11/90)
In article <1990Aug10.132443.23692@sbcs.sunysb.edu>, owen@sbcs.sunysb.edu (Owen Kaser) writes: |> You'd find that the main-frames are typically the best designs around, i.e. |> they will be the most conservative. They usually aren't done under the |> same constraints as designs at start ups, and they have more control |> over the parts supply quality. Even more to the point, you can get lucky with a 50 chip design. It is harder with a 500 chip design. With a 5000 chip design, if you don't design it worst case it will never work. I worked on large computers (multiple 30" racks of logic for a CPU) at a company that also built much smaller systems. The engineers that built the big systems thought that the small system guys did not do design in enough quality. The engineers that built the small systems thought that the big systems guys took too long in design and the resulting systems were way too expensive. Now, many years later, I see that both design methods were correct for their respective markets.