home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.arch
- Path: sparky!uunet!stanford.edu!snorkelwacker.mit.edu!bloom-picayune.mit.edu!athena.mit.edu!jfc
- From: jfc@athena.mit.edu (John F Carr)
- Subject: Re: DEC Alpha AXP System INTEGER Performance
- Message-ID: <1992Nov22.170912.26878@athena.mit.edu>
- Keywords: Alpha, AXP, SPEC, DEC, INTEGER
- Sender: news@athena.mit.edu (News system)
- Nntp-Posting-Host: achates.mit.edu
- Organization: Massachusetts Institute of Technology
- References: <1698@niktow.canisius.edu> <jdd.721687838@cdf.toronto.edu> <1992Nov20.220615.7494@raid.dell.com>
- Date: Sun, 22 Nov 1992 17:09:12 GMT
- Lines: 26
-
- In article <1992Nov20.220615.7494@raid.dell.com>
- samf@yosemite.dell.com (Sam Fuller) writes:
-
- >After reviewing the Alpha performance claims, I had the feeling that the
- >integer unit did not really run at 133Mhz. It looked to me that the
- >floating point pipeline was superpipelined at 133Mhz, but that the integer
- >side ran at a more reasonable 66Mhz. The pipes can operate independently
- >and I imagine be dispatched in parallel, hence the superscalar appelation.
-
- The integer pipeline runs at full speed, but:
-
- . dual issue is rare in the current implementation with current
- compilers, for integer-only code
-
- . many instructions, including load and shift, have multi-cycle
- latencies
-
- . the primary caches are small and the secondary cache has a
- long access time
-
- gcc in particular runs slowly on an Alpha, because it uses a lot of bit
- manipulation instructions, makes many memory references, and has a poor
- memory access pattern.
-
- --
- John Carr (jfc@athena.mit.edu)
-