home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.arch
- Path: sparky!uunet!europa.asd.contel.com!howland.reston.ans.net!usc!cs.utexas.edu!wotan.compaq.com!twisto.eng.hou.compaq.com!croatia.eng.hou.compaq.com!leigh
- From: leigh@croatia.eng.hou.compaq.com (Kevin Leigh)
- Subject: SCI vs. Anet (Compaq's I/O Proposal)
- Message-ID: <1993Jan11.225829.16388@twisto.eng.hou.compaq.com>
- Summary: Why didn't Compaq use SCI for I/O
- Keywords: SCI, I/O, point-to-point, Compaq, Anet
- Sender: news@twisto.eng.hou.compaq.com (Netnews Account)
- Organization: Compaq Computer Corp.
- Distribution: world
- Date: Mon, 11 Jan 1993 22:58:29 GMT
- Lines: 71
-
- ************************************************************
- We appreciate your suggestions given to us to consider SCI
- or one of it's derivatives (instead of Anet) as a solution
- for PC I/O interconnects. However, based on our
- interpretations of information we have, SCI and RamLink do
- not provide as low cost a solution as we desire. Here is
- why...
- ************************************************************
-
- "Many features are incorporated in SCI to support
- multiprocessor coherent and message-passing operations" [an
- excerpt from one of David Gustavson's SCI articles]. We
- agree.
-
- The problem that was addressed with Anet is much more
- limited in scope than the problem that SCI is intended to
- solve. All we ask of Anet is that it provides a mechanism
- for attaching I/O devices to the system. With this more
- limited scope, we were able to develop an interface that is
- very low cost while providing performance adequate for a
- wide range of I/O needs, up to and including graphics/video.
-
- Although much emphasis has been placed on low pin counts in
- all of our correspondence, getting low pin counts is not a
- sufficient cost metric. It has been explained to us by
- those knowedgeable about SCI that SCI implementations range
- from 50K to 100K gates in complexity. An Anet inteface can
- be done with <1K gates for a simple, slave-only device (e.g.
- a serial port controller) to around 10K gates for a multi-
- request mastering device (e.g. an FDDI controller.) This
- magnitude of difference is noticable no matter how
- inexpensive one can purchase ASICs.
-
- Another problem we have with SCI is its applicability in
- mobile applications. An SCI link is always "busy"
- transferring either information or idle packets (to
- synchronize the nodes), i.e., there is a constant power
- drain due to continuously switching devices. Athough Anet
- requires a clock, the total power in a quiescent Anet
- interface is much lower than one for SCI. To further reduce
- power consumption, the Anet protocol allows data buffers to
- be power gated when data is not actually being transferred
- and a further portocol allows the clock to be stopped to
- reduce power even more.
-
- It is interesting to note that one of the SCI design goals
- was Interface - "a standardized open communication
- architecture that allows products from multiple vendors to
- be incorporated into one system and interoperate smoothly"
- [IEEE Micro, Feb. 1992, pg. 11] - and yet there are a few
- different physical layers defined and more standards are on
- their way. We can see the usefulness of the parallel (P18)
- and the serial (S20) protocols for different distances. We
- can also see why RamLink (P1596.4) simplified SCI's original
- protocols significantly - primarily to interface to memory
- and not to worry about the MP issues. This is exactly what
- we were trying to point out - SCI has a lot of fat for low
- cost I/O applications and RamLink still is not lean enough.
- We tried to define Anet physical layer to be consistent
- across different performance systems. An Anet card should
- work in a very small computer as well as in a multiprocessor
- server.
-
- We actually view Anet and SCI as being complementry
- protocols just as we see both Anet and SCI as being
- complementry to existing inter-box protocols (e.g. Ethernet,
- FDDI, SCSI, etc.) We don't believe that there is a single
- connection mechanism that is optimal both for a multi-
- processing CPU/memory core and for I/O devices (especially
- I/O devices on something like a PDA.) So, different
- problem, different solution.
-