Path: utzoo!utgpu!water!watmath!clyde!rutgers!rochester!daemon From: miller@ACORN.CS.ROCHESTER.EDU (Brad Miller) Newsgroups: comp.lang.lisp Subject: Re: Troubles Message-ID: <6908@sol.ARPA> Date: 18 Feb 88 18:01:41 GMT Sender: daemon@cs.rochester.edu Lines: 75 Date: 17 Feb 88 04:40:09 GMT From: dzzr@beta.UUCP (Douglas J Roberts) 1) A Sun 3/260 w/16 MB will compare favorably to Symbolics' top-of-the-line 3670 in terms of run-time for large CL applications. A diskless 3/260 goes for about $35,800 (retail), whereas Symbolics is still in the one-full-up-machine-per-user mode, at about $75,000 per whack for a 3650 w/16 MB (slightly slower than the 3670). Slower? I thought it was about 1.4x *faster*. And a 16MB 3620 (the same processor, but slower disks, diffo cabinet) would only set you back about $55k. *NOT* diskless. If you are going to compare to a SUN, then add disks to the machine. Or count some fraction of the cost of your sever and remember that you are paging over the net. 2) The Sun 4/whatever: promises to be about about 3 - 4 times faster than the 3/260. (A guess - I haven't done any benches, but stuff presented in comp.arch would indicate the expected speed-up). By comparison, what new hardware does Symbolics have in the pipe? The only official thing we've heard about here is the Ivory VLSI chip machine. It will supposedly deliver a 3x runtime improvement over the 3670. And more, when geometries shrink. As if the growing hardware performance/$ gap were not sufficient reason for unease at Symbolics, Sun is also making some right moves WRT their LISP development environment. SPE, (Symbolic Programming Environment) promises to bring much of the LISP machine - like development and debugging functionality to the Sun's LISP environment. (The current Sun LISP development environment could kindly be described as "sparse"). Well, this too is an opinion, but from what I've heard, I'm not holding my breath. Lets face it, you can't make a lispm out of a box that is running UNIX. Now if you get rid of UNIX and have a unified address space sans kernel protections, if you can run >1 process that share *all* memory, if you can debug one process from another (no, ptrace is not sufficient), if you have incremental compilation of everything (including the kernel), if your machine understands objects instead of addresses, and a few dozen other small improvements, *then* maybe you have something. SPARC may be a strong step in this direction, but someone needs to port Genera to it. As a consumer of LISP hardware & software, I would hate to see Symbolics drop out of the running: competition is beneficial. On the other hand, it's a tough world out there... CP/M runs C. Lets say I give you a CP/M box just as fast as a SUN and it runs C faster than UNIX does. Would you buy it? I wouldn't. CP/M doesn't do what I need compared to UNIX, and never will without effectively *becoming* Unix. I can say the same about UNIX vs. Genera[*]. It just isn't sufficient. So, maybe this is the difference between buying a $20k pair of speakers and a 2k pair: is it 10x better? Nah, maybe 15%. Diminishing returns. But to some of us, it *is* worth it because our *time* is worth more than the incremental machine cost. I would hate to see Symbolics go, because it means we are all going to be stuck driving econobox chevy's --- with all those crowd pleasing features that have nothing to do with what *you* need. Thanks, but I don't *want* a machine that has been mass-produced to solve everybody's spreadsheet and array crunching problems. I want a machine that *helps me write software* that does things I didn't even know how to do when I started writing it. Brad Miller [*] Or TI's system. Maybe even Xerox's system: for the level of discussion in this article they are probably interchangeable. I have a Symbolics on my desk, so that's what I'm most familiar with. I don't mean to slight other lispm vendors. ------ miller@cs.rochester.edu {...allegra!rochester!miller} Brad Miller University of Rochester Computer Science Department 'If anything is True, this is.'