malcolm@Apple.COM (Malcolm Slaney) (07/24/90)
What are good machines for scientific visualizations? We have a Cray but I'm looking around for other machines that can do my work nearly as fast. My research (models of human hearing) can support parallelism and lots of vectorization. I want to be able to compute something and then display the result very quickly (20 frames per second) on a monitor. I guess the numbers that are important to me are >100 MFlops of peak performance and > 25 Mega-pixels per second output rate (at the same time.) Should I think about Convex? What about the new Connection Machine? Machines like the DAP and the MassPar seem to be too hard to program (I want to do new research in hearing, not algorithm development.) What do people think? Thanks. Malcolm Slaney Apple Perception Group malcolm@apple.com
rick@hanauma.Stanford.EDU (Richard Ottolini) (07/24/90)
In article <9811@hubcap.clemson.edu> malcolm@Apple.COM (Malcolm Slaney) writes: >What are good machines for scientific visualizations? >Should I think about Convex? What about the new Connection Machine? Machines >like the DAP and the MassPar seem to be too hard to program (I want to do >new research in hearing, not algorithm development.) No, Yes, ??, No. The two no's do not have integrated frame buffers or very high speed interfaces to frame buffers. The CM has one at the rate you specified. The MasPar is only slightly harder to program than the CM. Neither machine has made the layout of data in the memories fully transparent.
murray@sun13.scri.fsu.edu (John Murray) (07/25/90)
In article <9811@hubcap.clemson.edu> malcolm@Apple.COM (Malcolm Slaney) writes: >What are good machines for scientific visualizations? We have a Cray but >I'm looking around for other machines that can do my work nearly as fast. > >My research (models of human hearing) can support parallelism and lots of >vectorization. I want to be able to compute something and then display the >result very quickly (20 frames per second) on a monitor. It really depends on what sorts of visualizations you want to do. Do you just want to plop some big 2D scalar array on the screen? Or do you want 3D objects with smooth shading and lighting models? >I guess the numbers that are important to me are >100 MFlops of peak >performance and > 25 Mega-pixels per second output rate (at the same time.) > >Should I think about Convex? What about the new Connection Machine? Machines >like the DAP and the MassPar seem to be too hard to program (I want to do >new research in hearing, not algorithm development.) We have a Connection Machine CM-2. (64K processors, 256Kbits/proc) I have been having a wonderful time trying to do something useful with the graphics interface. The framebuffer data channel runs at 80Mbyte/second, but tuning a given program to get it to display anything at 20 frames/sec is a hefty challenge. Tuning a 3D rendering program to run at 20/sec might finally get me my masters... maybe a doctorate too :-) If you are doing complex graphics, you might be much better off with a current graphics work-station. Our Silicon Graphics IRIS-4D/240-GTX runs standard (not "tuned") programs at 10MFlop (real ones, not peak) per processor, and there are 4 processors on ours, can be up to 8 * . The graphics hardware processes about 40K smooth-shaded fully-lit polygons per second, again real ones - the salesmen say 100K lit, shaded polygons/sec. If this isn't enough, SGI's newest, the 300 series, has a 33% faster clock and much better graphics hardware. Up to eight processors, like the 200's. Also, there exist some mechanisms for remote displays - not just X-windows, fortunately - so with a fast enough data path you could be doing your math on a Cray and your 3D displays on an IRIS, real time. A final note on the CM-2: Thinking Machines IS working on greatly improving the graphics capabilities of the CM-2, but if you're not looking for a "Real Soon Now(tm)" solution... > Malcolm Slaney > Apple Perception Group > malcolm@apple.com * I have no idea if it's possible to parallelize something across multiple processors on the 4D series. I've never needed to try it. Disclaimer #1: I don't work for Silicon Graphics or TMC, I'm merely a satisfied user, etc., etc. Disclaimer #2: I do work for SCRI, but I sometimes wonder if they realize it. Meta-Disclaimer: "A society that needs disclaimers has too many lawyers" - Eric Pepke John R. Murray | "They call me Mr. Know-it-all, I am so eloquent. murray@vs2.scri.fsu.edu | Perfection is my middle name! | ...and whatever rhymes with 'eloquent'." - Primus
refson@castle.ed.ac.uk (Keith Refson) (07/27/90)
In article <9811@hubcap.clemson.edu> malcolm@Apple.COM (Malcolm Slaney) writes: >What are good machines for scientific visualizations? We have a Cray but >I'm looking around for other machines that can do my work nearly as fast. > >vectorization. I want to be able to compute something and then display the >result very quickly (20 frames per second) on a monitor. > >I guess the numbers that are important to me are >100 MFlops of peak >performance and > 25 Mega-pixels per second output rate (at the same time.) > >Should I think about Convex? What about the new Connection Machine? Machines >like the DAP and the MassPar seem to be too hard to program (I want to do >new research in hearing, not algorithm development.) As far as I know neither the Convex or DAP are visualization machines. I think you should consider machines from Stardent Computer Ltd (formerly Stellar and Ardent). They combine the sort of performance you seem to be seeking with the visualization. Our Stardent GS2000 can run my vector-parallel codes at 27% of the speed of a CRAY XMP (single processor) and do geometry rendering of complex molecular models at > 12 frames/sec. Codes vectorized for the cray should port easily, and the vectorizing fortran and C compilers are quite good. The greatest advantage I find of Stardent is their AVS visualization program which means that you can do vizualization without any graphics programming. At all. You should be able to integrate your codes with AVS with a small programming effort to get the computatational module to deliver data for immediate visualization. ---------------------------------------------------------------------------- | JANET : keith@uk.ac.ox.earth | Royal Mail: | | INTERNET: keith@earth.ox.ac.uk | Keith Refson | | BITNET : keith%uk.ac.ox.earth@ukacrl | Department of Earth Sciences | | UUCP : K.Refson@ed.uucp | Oxford University | | : keith%uk.ac.ox.earth@ukc.uucp | Parks Road | | PHONE : +44 865 272026/272016 | Oxford OX1 3PR | | FAX : +44 865 272072 | UK | ----------------------------------------------------------------------------