pkh@vap.vi.ri.cmu.edu (Ping Kang Hsiung) (09/22/89)
Does anyone have a collection of performance (timing) data based on running a raytracer (preferablly a publicly available raytracer, e.g. mtv or qrt) on various machines? In specific, I am interested in the answer(s) to the following question: To make a ray-tracing based animation film, which commercial machine(s) (excluding in-house hardware) can give the shortest computation time with moderate effort in optimization. For example, my experience indicates that mtv runs on a Cray Y-MP (-hintrinsic,o_level3) ~2.6x than on a PMAX, is this (Cray time) the fastest turn-around one can expect? Will a Connection Machine do better than this? What about an Intel iPSC? The Meiko transputer array? Silicon Graphics's Power series? Any info. is appreicated.
gavin@krypton.sgi.com (Gavin A. Bell) (09/22/89)
pkh@vap.vi.ri.cmu.edu (Ping Kang Hsiung) writes: >Does anyone have a collection of performance (timing) data based >on running a raytracer (preferablly a publicly available raytracer, >e.g. mtv or qrt) on various machines? I believe that the BRL-CAD ray-tracer is sometimes used as a standard benchmark (with specific input files). A number is generated which they call the 'Ray-tracing figure of merit'; the higher the number, the better. The whole BRL-CAD package is public domain, but big, crufty, and pretty ancient as ray-tracing packages go. I know all this because I'm in the Demo/Benchmarks group here at Silicon Graphics, and this ray-tracing benchmark is one of the few in which our 4D/280 outperformed a Cray (ray-tracing being an easily multi-processed, but not vectorizable, application). If people are intersted, I could send them the results. --gavin (Not an Official Spokesperson for anything)
spl@mcnc.org (Steve Lamont) (09/22/89)
In article <6224@pt.cs.cmu.edu> pkh@vap.vi.ri.cmu.edu (Ping Kang Hsiung) writes: >For example, my experience indicates that mtv runs on a >Cray Y-MP (-hintrinsic,o_level3) ~2.6x than on a PMAX, is this >(Cray time) the fastest turn-around one can expect? >Will a Connection Machine do better than this? What about an Intel iPSC? >The Meiko transputer array? Silicon Graphics's Power series? Well, it depends on what other things you feel prepared to do. I have both an IRIS 4D/120 and a Y-MP here (we just installed a Stardent Titan yesterday, also) and see the same sort of numbers. I'll run a couple of experiments today and get back to you with actual bench marks, if you'd like. I am currently working on a parallelized version of MTV and should be able to give you some results from that in a couple of days, barring getting blown away by Hurricane Hugo, that is :-). In any case, in order to take advantage of any parallelism on an MPP like a CM, you'll probably have to do a lot of recoding. The CM is a SIMD machine which does not really lend itself *directly* to ray tracing. I understand that TMC has done some interesting algorithmic development to actually *do* ray tracing, but the code certainly would not look anything like MTV as it stands right now. Barry, are you out there lurking? Any comments? The iPSC and Meiko systems would probably be more straightforward. If you're interested in results, contact me by private email and I'll be glad to share my thoughts with you. spl (the p stands for parallel) -- Steve Lamont, sciViGuy EMail: spl@ncsc.org NCSC, Box 12732, Research Triangle Park, NC 27709 "Surrealism only comes later when it seems 'reality' becomes difficult to achieve." - E. Miya, NASA Ames Research Center
phil@sem.BRL.MIL (Phil Dykstra <phil>) (09/26/89)
> The whole BRL-CAD package is public domain, ... The BRL-CAD package is *not* public domain. It is Copyright by the U.S.Army in order to control commercialization of it. We do distribute the source code at no charge however as long as the recipient agrees, in writing, to the conditions. > but big, crufty, and pretty ancient as ray-tracing packages go. Being one of the authors, I had to put in at least two cents worth of defense. Big - yes. Crufty - parts of it. Pretty ancient - some of it. But I wouldn't call things like CSG NURBs, arbitrary bounding planes, non-uniform space partitioning, parallel *and* network distributed capability very ancient. - Phil
badouel@irisa.irisa.fr (Didier Badouel) (09/26/89)
In article <6224@pt.cs.cmu.edu>, pkh@vap.vi.ri.cmu.edu (Ping Kang Hsiung) writes: > > Does anyone have a collection of performance (timing) data based > on running a raytracer (preferablly a publicly available raytracer, > e.g. mtv or qrt) on various machines? > We have a parallel raytracer (called PRay) at IRISA which as MTV uses NFF description databases. This raytracer has been implemented on an iPSC/2, on a SEQUENT BALANCE and also on serial computers (SUN3, GOULD NP1) to make better comparisons. I give you the various synthesis times for the well known 'Teapot' database. The image has been renderring with a 512X512 resolution with 3 light sources. Results are as follows : #PEs Time (in sec.) ________________________________________ SUN3: 8877 (~ 2h27mn) ________________________________________ GOULD NP1: 1642 (~ 27mn) ________________________________________ SEQUENT BALANCE 1 37121 (~ 10h18mn) 2 18567 3 12381 4 9285 5 7431 6 6197 7 5311 8 4656 9 4138 (~ 1h9mn) ________________________________________ iPSC/2 1 6294 (~ 1h45mn) 2 3332 4 1700 8 860 16 440 32 224 64 119 (~ 2mn) ________________________________________ The code running on the iPSC/2 emulates a virtual shared memory over the local PEs. The database is not duplicated but all the local memories are used. The remaining memory after loading code and data is used as a cache to speed up low global accesses. ________________________________________________________________ Didier BADOUEL badouel@irisa.fr INRIA / IRISA Phone : +33 99 36 20 00 Campus Universitaire de Beaulieu Fax : 99 38 38 32 35042 RENNES CEDEX - FRANCE Telex : UNIRISA 950 473F ________________________________________________________________