Path: utzoo!attcan!uunet!cs.utexas.edu!usc!zaphod.mps.ohio-state.edu!wuarchive!decwrl!shelby!cis!raje From: raje@dolores.Stanford.EDU (Prasad Raje) Newsgroups: comp.arch Subject: Re: Dynamic internal state (was Re: Scoreboards / Load-interlocks) Message-ID: Date: 19 Mar 90 22:34:22 GMT References: <11337@encore.Encore.COM> <3300106@m.cs.uiuc.edu> <32937@shemp.CS.UCLA.EDU> <132903@sun.Eng.Sun.COM> <39538@apple.Apple.COM> <22182@watdragon.waterloo.edu> Sender: news@cis.Stanford.EDU (USENET News System) Organization: Center for Integrated Systems, Stanford Lines: 48 In-reply-to: ccplumb@lion.waterloo.edu's message of 19 Mar 90 05:44:16 GMT > In article <39538@apple.Apple.COM> baum@apple.UUCP (Allen Baum) writes: > > If it's what I'm thinking of, its a cute idea. Guarantee that the 'valid' > > bit decays before any of the others, and its easy. > > That is *sneaky*. Now, refresh the entry in case of a cache hit. If > the entry hasn't been used in a few milliseconds, it will be purged. > Not-recently-used! It may be impractical for other reasons, but it has > ^^^^^^^^^^^^^ > serious hack value. > -- > -Colin ( The suggestion is to use dynamic storage for the valid bit and rely on its decay to purge unused entries in the cache ) Unfortunately this wont work, for this "other" reason. The rate of decay of a DRAM memory cell (essentially the charge leakage from the storage node through a reverse biased p-n junction) is very difficult to control accurately. After the junction leakage has been reduced to the order of a few pico amps per micron^2 it is very difficult to control to even 100% accuracy. The refresh rate of the order of a few milliseconds is really a safe lower limit on how long the charge will reliably be stored on the node. There have been recent DRAMs made with as long as 1 second retention times. Assuming a storage capacitor of C = 50 fF and a leakage current of i = 10 pA using i = C dv/dt, the rate of decay is 0.2 V/ms So even if you were able to control i very accurately to 10pA, the cell voltage have decayed only to 4.8V from 5V after 1ms. When the time comes to decide if the valid bit has "faded away" it will be very difficult to make a decision in a time that is of the order of sucessive cache accesses. In brief, a highly uncontrollable technology-level parameter would determine a architecture-level parameter viz. the replacement strategy of the cache. I realize this is straying away from comp.arch, but this thread has been here for a while now. cheers Prasad