wayne@pur-phy (Wayne K. Schroll) (02/02/89)
I have heard that the 68882 is pin-equivalent to the 68881. Is this
true? If so, can I just pop the '881 out of my Mac II (A/UX) and drop
in the '882? I have also heard that the '882 will provide a speed
increase of about a factor of two on floating point operations. Is this
true? Is there anything special I should know since my Mac II runs A/UX?
Finally, does anyone know of a good place to purchase a 68882 chip, and
it's approximate cost? Thanks for any information you can provide.
Wayne
------------------------------------------------------------------------------
Wayne K. Schroll --* --* --* -* wayne@newton.physics.purdue.edu
Department of Physics -* ---* -* -* --* --* -* ---* -*
Purdue University -* -* -* ---* ---* --* !pur-ee!pur-phy!wayne
------------------------------------------------------------------------------schmitz@fas.ri.cmu.edu (Donald Schmitz) (02/04/89)
In article <1931@pur-phy> wayne@pur-phy (Wayne K. Schroll) writes: > > I have heard that the 68882 is pin-equivalent to the 68881. Is this >true? Yes. > If so, can I just pop the '881 out of my Mac II (A/UX) and drop >in the '882? Maybe. The '882 generates ~34 bytes (worst case, can't remember exact number) more state information on an fsave than an '881. If the operating system allowed extra space for this in the process control block (you must save this state on every context switch) everything will work fine. Some operating systems, SUN OS 3.X in particular, did not do this, and strange things will likely happen if you pop in an '882. SUN warned us not to try this, and our system manager agreed. We did install an '882 in a standalone system with a locally written real-time kernel (I wrote the context switch code, and was happy I left an extra 128 bytes in the save area).I'd be interested in hearing if Apple did this right. > I have also heard that the '882 will provide a speed increase of about a > factor of two on floating point operations. Is this true? Yes. The big advantage of the '882 is hardware assisted (combinational logic) fmove. fmove is much faster than the '881, and with typical FP code being half fmoves, most code will go twice as fast without recompilation. If you want to get down to the assembler level, you can do even better. The '882 allows an fmove to execute in parallel with an arithmetic operation, as long as both do not use the same FP registers. By carefully organizing code, you can triple the speed of an '881. This is only practical for small pieces of code. >Finally, does anyone know of a good place to purchase a 68882 chip, and >it's approximate cost? I bought our early sample through Pioneer Electronics, our local Motorola distributor, I think they are a national chain. Cost for a 16 MHz part, quantity 1, was ~$400, but that was some time ago. Don Schmitz schmitz@fas.ri.cmu.edu --