Article View: comp.arch
Article #1608Re: Need 50Mflops on a VME board
From: Keith Bierman - SPD Languages Ma...
Date: Thu, 20 Jul 1989 21:11
Date: Thu, 20 Jul 1989 21:11
42 lines
2708 bytes
2708 bytes
In article <5583@pt.cs.cmu.edu> ralphw@ius3.ius.cs.cmu.edu (Ralph Hyre) writes: >Matrix multiplication is the critical operation, and we could probably live >with 2 25Mflop boards, as long as we could effectively use them in parallel. >We'll be feeding them with a Sun-3/280 or SparcServer 330, so they need >to be 6UVME (preferred) or 9U boards that could coexist with the Sun memory >bus compatible. > >I'm famimiliar with a Weitek-based board from Mercury, and there seem >to be several DSP-based solutions out there (are 96000-boards avilable >yet?) single precision and maybe even fixed point seem to be OK. > >Is such a beast obtainable for around $50K? I expect that the VMEbus >performance of the Sun may be the limiting factor. > >-- > - Ralph W. Hyre, Jr. I have used the mercury board in a sun3/160 and a sun4/260 (I have not moved it to my sun 4/330 yet). It functions well; the biggest limiting factor in performance is its own memory system ... I do not think that matrix multiply happens at more about 8mflops. The 20+mflop peak is only approached for their fft code (meaning that perhaps matmult can be cleverly recoded). Mercury provides a GreenHills (-gag-<) derived compiler which can run your entire code on the board (one can get at the knobs, and run multiple boards, and etc.) for applications which lend themselves to this, there may not be an advantage in having a 4/330. I believe that they have several customers whose applications have been coded to run on multiple boards to achieve speedup. They do not (yet ?) have "microtasking", so its all done by hand. The current mercury board is really 6U with an adapter. It is rumored that a soon to be released version might be really 9U and have 64Mb of its own local RAM. I am not hip to prices. cheers Keith H. Bierman |*My thoughts are my own. Only my work belongs to Sun* It's Not My Fault | Marketing Technical Specialist ! kbierman@sun.com I Voted for Bill & | Languages and Performance Tools. Opus (* strange as it may seem, I do more engineering now *)
Technical Details
Message-ID:
<116825@sun.Eng.Sun.COM>Path:
arc1-ca.pugleaf.net!utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!bionet!ames!apple!sun-barr!newstop!sun!chiba!khbReferences:
<626@megatek.UUCP> <115695@sun.Eng.Sun.COM> <5583@pt.cs.cmu.edu>Show Full Headers
Complete Article Headers
Path: arc1-ca.pugleaf.net!utzoo!attcan!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!bionet!ames!apple!sun-barr!newstop!sun!chiba!khb From: khb%chiba@Sun.COM (Keith Bierman - SPD Languages Marketing -- MTS) Newsgroups: comp.arch Subject: Re: Need 50Mflops on a VME board Message-ID: <116825@sun.Eng.Sun.COM> Date: 20 Jul 89 21:11:11 GMT References: <626@megatek.UUCP> <115695@sun.Eng.Sun.COM> <5583@pt.cs.cmu.edu> Sender: news@sun.Eng.Sun.COM Reply-To: khb@sun.UUCP (Keith Bierman - SPD Languages Marketing -- MTS) Distribution: na Organization: Sun Microsystems, Mountain View Lines: 42