pls@ncsu.UUCP (Phillip L. Shaffer) (03/06/85)
I am trying to find out execution times for floating-point instructions for a DEC PRO-350. This is needed for analysis of some experiments we conducted to compared predicted and actual execution times of some signal-processing type algorithms on microprocessors. So far, DEC has been unhelpful. It appears that the PRO-350 has a KEF11-CA microcode option that adds some 46 f.p. instructions to the integer instruction set, rather than the FPF-11 f.p. co-processor. I found some times listed for KEF11A microcode option in the 1979-80 DEC Microcomputer Processor Handbook, but am suspicious of these old figures: ADDD add double 42.45 microsec SUBD subtract double 43.35 microsec MULD multiply double 193.05 microsec DIVD divide double 239.25 microsec. There appears to be no logarithm instruction (would like times if there is one). I would appreciate any data or pointers on this (like: who knows at DEC, what manual contains this info, etc.). Related to this, we ran the same set of programs on the DEC PRO-350 and on an IBM PC/XT, and have total times for both: for both, the programs were written in C, and were identical. The PRO was running Venix, the PC/XT (with 8087) was running PC-DOS 2.02 and we used the DeSmet C compiler. We were surprised that the PC/XT times were 20-50% FASTER than the PRO! This seemed particularly odd, as the DeSmet compiler generates rather inefficient 8087 code (subroutine calls for each operation). Any comments on this? Thanks for any help anyone can give. Phil Shaffer P.S. I just talked with a "hardware support specialist" from DEC, and he said (1) the PRO-350 has a FP-11 coprocessor (KEF-11-CA is the configuration number); (2) "the execution time is in the range of 1.11 MHz to 5 MHz depending on the function of the CPU." He said he couldn't be more specific; he obviously didn't know what he was talking about. Is anybody knowledgeable from DEC out there?
michael@nvuxd.UUCP (M.CAIN) (03/12/85)
The exact nature of the benchmark can, of course, influence the results. We've tried some floating point signal processing routines, and the Pro350 and PC/XT (both running VENIX) were pretty close (on the order of 5-10%). I suspect that this was so because the routines were manipulating large arrays of floating point numbers, and the PC was slowed substantially by its 8-bit bus. I seem to recall seeing at least one "benchmark" that showed the PC with 8087 was about 0.60 VAX, but used so few floating point values that they could all be kept in the 8087 registers. Michael Cain Bell Communications Research ..!bellcore!nvuxd!michael
wjafyfe@watmath.UUCP (Andy Fyfe) (03/15/85)
There are a number of DEC PRO 350s around the campus here. A number of
them are used by the applied math department for various numerical things
(with many of the programs written in basic). We found out that the
basic square root function is hopelessly inaccurate -- no more than
about 3 significant digits. Raising to the power .5 is much better.
Is this the fault of basic, or the floating point hardware?
--Andy Fyfe ...!{decvax, allegra, ihnp4, et. al}!watmath!wjafyfe
wjafyfe@waterloo.csnetdan@rna.UUCP (03/15/85)
x You are basically right. The Pro 350 uses the F-11 chip set. The KEF-11 chip is a floating point microcode extension. The time you quote for floating point performance are for the 11/23 with the KEF-11 but should be close to correct. DEC also makes a bit-slice FPF-11 which is plug compatible with the KEF-11 chip and is said to be 5X faster than the KEF-11 chip. You should be able to use that board if you provide it with power (it normally obtains power from either the Qbus or Unibus slot). There is no logarithm instruction (or any transcendentals) on any PDP-11 machine. My benchmarks on an 11/23 and a PC/XT (amongst other machines) were posted a while ago. For floating point performance, the 11/23 rated at .034 of an 11/780 (w/ FPA) while the PC/XT rated at .13, i.e. the PC/XT should be as much as 3-4X faster than 11/23 (or the PRO350) in floating point. Overall, however, the 11/23 rated at .16 of the 11/780 while the PC/XT rated at .12 . The tested 11/23 had a fast disk. I further understand that the PRO350 disk is abysmally slow, so it would not rate so high. Incidentally, the 11/73 floating point rates at .16, overall .35, so the PRO380 should be more reasonable since it also uses the J-11 chip.