home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.sys.intel:2670 comp.arch:11605
- Newsgroups: comp.sys.intel,comp.arch
- Path: sparky!uunet!spool.mu.edu!umn.edu!csus.edu!netcom.com!norm
- From: norm@netcom.com (Norman Hardy)
- Subject: Re: Superscalar vs. multiple CPUs ?
- Message-ID: <1992Dec13.063930.8687@netcom.com>
- Organization: Netcom - Online Communication Services (408 241-9760 guest)
- References: <1992Dec10.173032.1418@twisto.eng.hou.compaq.com> <1992Dec11.060038.27494@netcom.com> <PRENER.92Dec12010123@prener.watson.ibm.com>
- Date: Sun, 13 Dec 1992 06:39:30 GMT
- Lines: 30
-
- In article <PRENER.92Dec12010123@prener.watson.ibm.com> prener@watson.ibm.com (Dan Prener) writes:
- >In article <1992Dec11.060038.27494@netcom.com> norm@netcom.com (Norman Hardy) writes:
- >
- >> How about sharing expensive functional units (multiply, divide,
- >> DES) between processors on the same chip. This gives the economic
- >> advantages of heterogeneous processors (division of labor) without
- >> the software pain. The processors would appear to the operating
- >> system as CPUs sharing memory. Indeed this approach makes special
- >> purpose hardware more attractive as functional units.
- >
- >Doesn't that make saving the state of a processor somewhat complicated?
- >--
- > Dan Prener (prener@watson.ibm.com)
-
- I see no need to change the programmer's model in order to share
- functional units. Functional units have no state if you interrupt
- between instructions as in most architectures. On processors such
- as the 88K the floating point units are frozen upon interrupt so
- that interrupt latency may be less than the divide latency. At
- worst this can be handled by providing a register per processor
- per functional unit to capture the results of operations issued
- by the interrupted processor. The interrupt hardware might even
- be cagey and favor for interruption a processor with no
- outstanding instructions in shared units.
-
- The 88K has a systolic multiplier and I imagine that few programs
- issue multiples more than 10% of the cycles. I imagine that
- sharing such a multiplier between four instruction streams would
- produce little interference. Perhaps the 88K has too much
- multiply hardware. I don't know the tradeoffs.
-