modern stuff

On Fri, 26 Oct 2018, emanuel stiebler via cctalk wrote: I worked for a company that made supercomputer boards out of the i860 at one point - at the time (very early 90's) they were blindingly fast, 40Mhz, 3 instructions per clock cycle which, since one was a floating point multiply and add meant that it was pretty good - at the time. However it was a royal PITA to code for although a 32-bit CPU, it would read memory 64 bits at a time (actually 128 IIRC to satisfy the cache), with half that 64-bit word being an instruction for the integer unit and half for the floating point unit, so you effectively had to build a floating point pipeline by hand coded instructions, so 8 (I think) instructions to load the pipeline, then each subsequent instruction would feed another value into the pipe, then another 8 at the end to empty it. Great for big matrix operations, rubbish for a single add of 2 FP numbers. The issue came when you wanted to take an interrupt - the overhead of flushing the pipe, reloading it all for the next context, and so on really bogged it down. Not to mention writing assembly code in 2 columns... There were quite a few systems built with about 30 boards in them, each having 2 x i860's and a good few MB of RAM (64MB I think) built. -Gordon