gnu@sun.uucp (John Gilmore) (02/26/85)
Firstly I should point out that what we are really talking about is 24 or 32 bits of VIRTUAL address space, not physical. You can have as few or as many bits on the other side of the MMU as you like; what comes out of the (320xx or 680xx) CPU chip is a virtual address. Now, Chuqui sez: > We DO have 32bit addressing > silicon on the boards and on its way to reality, and I expect it will be a > purchasable commodity long before the megabit chips that will be neccessary > to make those kind of address spaces truly useful. AT&T is advertising that they will have production 1MB rams for sale before the end of 1985. (Of course, they haven't mentioned price, pinouts, or anything else...and have never sold chips on the market before.) 32bit addressing makes a > nice marketing tool, granted, but there really isn't much that a 32bit > address gives you that a 24bit address doesn't also give you in a > manufactured product EXCEPT a marketing tool. There are a few kinds of applications that really need more than 16 megabytes of address space. Large, highly correlated databases are the obvious example, e.g. CAD simulation, linear programming, heavy image processing. A friend who consults to Intel reports that their production scheduling system has hit the limits of the IBM 370's 16-meg address space -- mostly due to misdesign of the software -- and requires constant tweaking, or a complete redesign, to make it run. I was hearing similar complaints (I don't remember who from) about large electronic simulations 3 to 4 years ago. This is not to say that a 16MB virtual space is very limiting. It's not. Our systems provide the full 16MB per process on the 68010, and nobody uses it. (They don't want to allocate that much swap space anyway!) It's just that when you need it, you really need it, since recoding to deal with the limit will take months to years and/or slow down the (already slow, since it's dealing with 16MB of data) program by a good factor. The people who need that capability soon will buy Motorola. National can certainly afford to ignore that small segment of the market for the next year or two. In other words, Chuqui is right...
sean@ukma.UUCP (Sean Casey) (03/01/85)
Special CAD controllers and extremely large databases
are examples of a rare exception, rather than the rule.
Attacking National for only providing 24 bits of virtual
address is not only unfair, it is completely unreasonable.
People that use that kind of memory also need specialized
hardware to give the speed they require. If I was looking
for a CAD or Dbase system that needed > 16MB, you can be
damn sure I'm not going to go with any generic minicomputer
processor such as the 32032 or 68000. I'm going to need FAST
memory, and FAST dedicated graphics and array processors if
I don't want to wait years for results. The 68xxx and the
32032 just don't have the MIPS for that sort of work. I'd be
looking at large mainframes or dedicated workstations first.
You can find faults with any processor you look at. The
32032 is a milestone in design, not only for the MMU, but
also for it's rich and higher language oriented instruction
set. Cutting it down is like cutting down a Porsche 924 for
not being a Ferrari Boxer. Certainly it's not. But look how
much the other costs.
Sean
--
Sean Casey UUCP: {hasmed, cbosgd}-\
{ucbvax, unmvax, boulder, research}!anlams---ukma!sean
{mcvax!qtlon, vax135, mddc}!qusavx-/
ARPA: "ukma!sean"@ANL-MCS or sean%ukma.uucp@anl-mcs.arpa