home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.arch
- Path: sparky!uunet!newsgate.watson.ibm.com!yktnews!admin!florida!essbaum
- From: essbaum@rchland.vnet.ibm.com (Alexander Essbaum)
- Subject: Re: Alpha and Super-pipelining
- Sender: news@rchland.ibm.com
- Message-ID: <1992Dec30.200629.13254@rchland.ibm.com>
- Date: Wed, 30 Dec 1992 20:06:29 GMT
- Disclaimer: This posting represents the poster's views, not necessarily those of IBM
- References: <1992Dec30.175717.15249@dcc.uchile.cl>
- Nntp-Posting-Host: florida.rchland.ibm.com
- Organization: IBM Rochester
- Lines: 45
-
- In article <1992Dec30.175717.15249@dcc.uchile.cl>, pchris@dcc.uchile.cl (Chris Perleberg) writes:
- |>
- |> After reading the material available about the Alpha, I noticed that it was
- |> mentioned a few times that the first implementation of the Alpha, the 21064,
- |> is *super-pipelined* as well as superscalar. For example, in "infosheet.txt"
- |> that is available on gatekeeper.dec.com the following is stated:
- |>
- |> > Digital's 21064 Microprocessor
- |> > ....
- |> > The 21064 is a super-scalar, super-pipelined implementation of the
- |> > Alpha architecture. Super-pipelined means that an instruction is issued
- |> > to the functional units at every clock tick and the results are
- |> > pipelined. Being super-scalar, the architecture allows the instruction
- |> > unit to issue two instructions per clock tick, resulting in
- |> > significantly higher throughput and performance.
- |>
- |> After reading the Data Sheet of the 21064, I find nothing that seems to
- |> indicate the 21064 is super-pipelined, at least in terms of the common
- |> definition of super-pipelined. Is DEC redefining the word "super-pipeline"?
- |> Anybody out there that can explain?
- |>
- |> Chris Perleberg
- |> pchris@dcc.uchile.cl
-
- i always thought super-piplining referred to piplining funtions that would
- normally be given a single pipeline stage. ie. if a processor has 4 stages
- with delays like:
-
- ifetch 10ns
- decode/operand fetch 10ns
- execute 20ns
- writeback 10ns
-
- you can cut the critical path delay by 10ns if you break the execute into
- 2 stages:
-
- ifetch 10ns
- decode/operand fetch 10ns
- execute1 10ns
- execute2 10ns
- writeback 10ns
-
- the second example is super-piplined.
-
- axel
-