home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.lsi:542 comp.lsi.cad:879
- Path: sparky!uunet!wupost!gumby!yale!mintaka.lcs.mit.edu!ai-lab!life!tk
- From: tk@ai.mit.edu (Tom Knight)
- Newsgroups: comp.lsi,comp.lsi.cad
- Subject: Re: Multiported Register Files
- Message-ID: <TK.92Aug17213322@wheat-chex.ai.mit.edu>
- Date: 18 Aug 92 01:33:22 GMT
- References: <2A8C14C1.24975@ics.uci.edu>
- Sender: news@ai.mit.edu
- Followup-To: comp.lsi
- Organization: MIT Artificial Intelligence Lab
- Lines: 37
- In-reply-to: andrea@esp.ics.uci.edu's message of 14 Aug 92 20:17:06 GMT
-
- In article <2A8C14C1.24975@ics.uci.edu> andrea@esp.ics.uci.edu (Andrea Capitanio) writes:
-
- From: andrea@esp.ics.uci.edu (Andrea Capitanio)
- Summary: Do anybody knows anything about multiported SRAMs with more than 6 ports ?
-
- I'm wondering if anybody out there is aware of any implementation
- of multi-port static memories with large (i.e., >= 6) # of ports
- or if any reference to the topic exist.
-
- There are a few ideas that are "well known" that should be mentioned.
- One is that you can duplicate the register file any number of times to
- make multiple read ports. The real problem is multiple write ports.
- A nice arrangement is to share write bit lines across a pair of
- arrays, and split read bit lines between the two arrays. Then every
- write is seen by both arrays, and there are independent read ports on
- each side of the pair of arrays.
-
- IBM used single ended reads and writes in the RT processor. If you
- know that you will do both a read and a write in a single cycle, then
- shared bit lines aren't a plus. You can specialize the bit line
- wiring to distinguish read and write bit lines -- the read lines can
- get driven by large devices, and the feedback inverter, which must be
- overpowered by the write data, can be made small. This was in
- in the technical paper collection IBM published on the RT in about
- 1987. I'm not sure if it ever made it into a journal. "Romp/MMU
- Circuit Technology and Chip Design" R.A. DuPont, et al., p. 86, in IBM
- RT Personal Computer Technology, Form SA23-1057, IBM Engineering
- Systems Products, 472 Wheelers Farms Rd., Milford, CT 06460.
-
- You might be amused (but probably only amused) by the paper of Rivest
- and Glasser "A Fast Multiport Memory Based on Single Port Memory
- Cells" in which they propose bulding multiport rams which handle
- addressing conflicts by intentionally introducing errors into the
- array, but only in single bit positions, such that the errors can be
- later corrected on a read operation. MIT Lab for Computer Science
- memo TM-455 July 1991.
-
-