Article View: comp.sys.cdc
Article #144Re: Cyber Speeds
From: Patrick F. McGehearty
Date: Fri, 08 Jan 1993 16:53
Date: Fri, 08 Jan 1993 16:53
44 lines
2994 bytes
2994 bytes
In article <martineau-080193081105@macmartineau.ccr.hydro.qc.ca> martineau@MacMartineau.ccr.hydro.qc.ca (Alain Martineau) writes: ...in response to request for info about CDC Cybers vs 486PCs... > >The LINPACK performance table can be obtained by sending to netlib@ornl.org >a message saying: >send performance from benchmark > >Here is what I found: >CDC Cyber 2000V 32 Mflops >CDC Cyber 205(4-pipe) 17 Mflops >CDC Cyber 205(2-pipe) 17 Mflops >CDC Cyber 990E 12 Mflops >CDC Cyber 4680 9.4 Mflops >CDC Cyber 875 4.8 Mflops >CDC Cyber 176 4.6 Mflops >CDC Cyber 760 2.6 Mflops >CDC Cyber 175 2.1 Mflops >CDC Cyber 180-860 2.1 Mflops > >The first 486 listed is >Compaq Deskpro 486/331-120w/487 1.4 Mflops > As a further clarification for those not familiar with the Linpack benchmark suite, these numbers reference the 100x100 version of Linpack, which requires an all Fortran solution. The 1000x1000 version allows architecture specific algorithms and assembly language tuning. The differences can be substantial. For example, the Cyber 205(4-pipe) is listed at 195 Mflops for the 1000x1000 problem set vs 17 Mflops for 100x100 problem set, an improvement of over 1000%. The range of difference varies strongly with architecture, with vector and parallel architectures typically showing much larger gains for the larger problem than scalar architectures. For example, the DEC VAX 6000/410 (1 processor, scalar) shows 1.2 Mflops on 100x100, and only 1.5 Mflops on 1000x1000, an improvement of only 25%. Vendors who do not expect significant gains on the 1000x1000 problem over the 100x100 problem are inclined to not submit results for the 1000x1000 problem. So, my conclusion is that the linpack comparison shown above tends to understate the floating point performance advantage of the Cybers over a PC. Readers should also be aware that Linpack performance says nothing about integer or I/O performance. If this discussion whets your appetite for more architecture/performance comparisions, read comp.benchmarks for a few weeks. Also be prepared for the constant cavet 'your mileage will vary'.
Technical Details
Message-ID:
<1993Jan8.165340.8396@news.eng.convex.com>Path:
arc1-ca.pugleaf.net!sparky!uunet!cs.utexas.edu!hermes.chpc.utexas.edu!news.utdallas.edu!convex!patrickReferences:
<1993Jan07.115610.21080@crash> <martineau-080193081105@macmartineau.ccr.hydro.qc.ca>Show Full Headers
Complete Article Headers
Newsgroups: comp.sys.cdc
Path: arc1-ca.pugleaf.net!sparky!uunet!cs.utexas.edu!hermes.chpc.utexas.edu!news.utdallas.edu!convex!patrick
From: Patrick F. McGehearty <patrick@convex.COM>
Subject: Re: Cyber Speeds
Originator: patrick@wagner.convex.com
Sender: usenet@news.eng.convex.com (news access account)
Message-ID: <1993Jan8.165340.8396@news.eng.convex.com>
Date: Fri, 8 Jan 1993 16:53:40 GMT
Reply-To: patrick@convex.COM (Patrick F. McGehearty)
References: <1993Jan07.115610.21080@crash> <martineau-080193081105@macmartineau.ccr.hydro.qc.ca>
Nntp-Posting-Host: wagner.convex.com
Organization: Engineering, CONVEX Computer Corp., Richardson, Tx., USA
X-Disclaimer: This message was written by a user at CONVEX Computer
Corp. The opinions expressed are those of the user and
not necessarily those of CONVEX.
Lines: 44