home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!haven.umd.edu!decuac!pa.dec.com!e2big.mko.dec.com!nntpd.lkg.dec.com!sousa.tay.dec.com!human.enet.dec.com!supnik
- From: supnik@human.enet.dec.com (Bob Supnik)
- Newsgroups: comp.arch
- Subject: Re: uniprocessor design ceiling
- Summary: Trend can continue for w while
- Keywords: microprocessor, evolution, performance
- Message-ID: <2340@sousa.tay.dec.com>
- Date: 26 Dec 92 19:22:05 GMT
- References: <1gbpjrINNsq6@girtab.usc.edu>
- Sender: newsa@sousa.tay.dec.com
- Organization: Digital Equipment Corporation
- Lines: 61
-
-
- In article <1gbpjrINNsq6@girtab.usc.edu>, plaw@usc.edu (Patrick Law) writes...
- >Correct me if I am wrong. For the last 20 years, we have 4-bit
- >microprocessor, CISC, cache, RISC, superscalar, superpipeline,
- >etc. Have we reached a point of diminishing return in uniprocessor
- >design? Could we still get a speedup of "X" every "N" years like
- >before? <text deleted>
-
- Since the mid-70's, microprocessors have exhibited performance growth of
- roughly 50% per year, within a given base architecture type (eg, CISC
- processors, RISC processors, superscalar RISC processors). Since circuit
- speed has not improved as much (roughly 30% every process generation),
- the balance has come by improved design and microarchitecture. An example
- from VAX history:
-
- chip year process cycle rel comments
- perf
-
- MicroVAX 1985 3u 200ns 1X No pipeline within microcycle
-
- CVAX 1987 2u 80ns 3X 'Half folded' pipeline within
- microcycle (ie, register writes
- overlapped with reads), onchip cache
-
- CVAX+ 1988 1.5u 60ns 4X Straight shrink
-
- Rigel 1989 1.5u 28ns 8X 'Fully folded' pipeline within
- microcycle (register read/ALU/register
- write overlapped, like classic RISC)
-
- Mariah 1990 1.0u 16ns 14X Straight shrink + circuit tweaks
-
- NVAX 1992 .75u 12ns 32X Overlapped macropipeline -> lower cpi
-
- In RISC, a comparable timeline would show the five stage pipe 'classic RISC'
- employing more finely divided pipelines (for faster clock rates) and/or
- multiple instruction issue, as well as improvements in memory hierarchy.
-
- The ingenuity of processor designers (and semiconductor technologists) being
- far from exhausted, I believe this performance growth can continue for the
- rest of this decade. There are some limiting factors:
-
- - Process technology evolution appears to be slowing somewhat, due to
- lengthening development time (and cost) for new generations of semiconductor
- processing equipment.
-
- - Increases in performance have been accompanied by increases in power
- consumption, in opposition to rising requirements for 'green' technology.
-
- The shape of microprocessors in the year 2000 is the subject of a panel
- discussion at the ISSCC conference, February 93.
-
- Finally, the ability to continue improving uniprocessor performance does
- negate the importance of cracking the programming problems on MIMD's and
- other collectives. An MPP could provide one to three orders of magnitude
- more computing power. Indeed, if the underlying microprocessors continue
- to improve, so will the MPP.
-
- Bob Supnik >Supnik@human.enet.dec.com
- >All opinions expressed are those of a hardline microcoder
- >and do not reflect those of Digital Equipment Corporation
-