ughtas@diablo.amd.com (Tom Stewart) (08/04/90)
Now that I've moved up to the latest revision of rayshade, things are wonderful :-). It also runs rings around the previous version. I've taken to running my 29K coprocessor at work almost 24 hrs. a day cranking images out. The down side is that I'm now forced to wait hours, days, or even weeks for images on my AT clone at home :-(. I'm thinking about getting an IIT 2C87-20 fpu that's become available recently. Apparently, this is a wazoo new implementation of 287/387 FPU technology in a '287 package. The ads I've seen claim up to 20x over a standard 80287 in an AT. They also claim 24 extra regs and matrix instructions. Obviously one would need assembler routines to use these features, but how much would the features actually help out? Any comments? [ or are all you all to busy soaking up CPU-hrs on your CRAY's ?? :-) ] Tom Stewart
ganter@urz.unibas.ch (08/05/90)
In article <1990Aug3.185628.29023@mozart.amd.com>, ughtas@diablo.amd.com (Tom Stewart) writes: > Now that I've moved up to the latest revision of rayshade, things > are wonderful :-). It also runs rings around the previous version. > I've taken to running my 29K coprocessor at work almost 24 hrs. a > day cranking images out. > > The down side is that I'm now forced to wait hours, days, or even > weeks for images on my AT clone at home :-(. > > I'm thinking about getting an IIT 2C87-20 fpu that's become available > recently. Apparently, this is a wazoo new implementation of 287/387 > FPU technology in a '287 package. The ads I've seen claim up to 20x > over a standard 80287 in an AT. They also claim 24 extra regs and > matrix instructions. Obviously one would need assembler routines > to use these features, but how much would the features actually help out? > > > Any comments? [ or are all you all to busy soaking up CPU-hrs on your > CRAY's ?? :-) ] > > > Tom Stewart I don't think, that the IIT runs 20x the good old 287/387 but 20% more speed is not exagerated, because of better architecture and specially (interesting for slow ATs) the turbo-socket with all glue parts needed to run the FPU asynchronically with faster speed (on a 12MHz 286 a 20 MHz 2C87 can be optimally used, if floating-point routines are optimized). The 24 extra registers ? Well, as I said, if You optimize code, You can load registers in advance and then do all hard work without lot of lost time in changing registers. The main reason of more registers are the matrix instructions. You can't do it as fast with conventional algorithms using multiple instructions. These assembler routines are written very fast, it's worth the effort. And one point more: because the IIT is faster, it is cheaper... (really). Robert