oconnor@sunset.steinmetz (Dennis M. O'Connor) (02/20/88)
In re the how-big-an-address-space-do-you-need question : Well, for PHYSICAL space, I am reminded of yet another "Amdahl's Law": X "MIPS" == X Megabytes == X Mbits/sec I/O. The opinion is that this indicates a computer system well-balanced for general-purpose computation. It assumes virtual memory space in excess of the physical memory space, and that the paging device is part of that high-speed I/O and does NOT have a high latency ( like paging over Ethernet appears to sometimes have ). Since what a "MIPS" is is fuzzy, obviously the "identity" above will be fuzzy as well. But it seems to be within a factor of 4. So, SUN-4s should have (@7"MIPS") 8 to 32 MBytes of RAM. A Cray XMP (@800"MIPS"?) should have maybe 800 to 3200 MBytes. But what about Symbolics LISP Machines ? These claim to be 1 MIP workstations ( the 3600s anyway ) but are NOT very happy with only 1-4 MBytes of RAM. Well, the "cop-out" is that they are NOT "general purpose" computing. Is a Cray ? Signal Processors are also NOT general purpose : they are happy with a good bit less than 1MByte/MIPS. My feeling on the matter is that as CPU performance climbs, RAM needs will climb as well. We'll need more than 32-bits of address space for 1000MIPS machines. If I was designing a microprocessor architecture today, ( I WAS designing one last year, but that's finished ) I wouldn't let it worry me. Micros have already hit 40MIPS, and will hit 100MIPS in the next five years, but (fearless prediction ahead) MICRO-processors wont hit 1000MIPS until the year 2005 or so, if even then. How many microprocessor architectures from 1965 are still around today? :-) As for virtual address size : wellb beats me. Seems to me virtual memory is more like closet space than physical memory is. If you've got it, you'll use it. -- Dennis O'Connor oconnor@sunset.steinmetz.UUCP ?? ARPA: OCONNORDM@ge-crd.arpa "Nuclear War is NOT the worst thing people can do to this planet."
pf@diab.UUCP (Per Fogelstr|m) (02/26/88)
In article <9629@steinmetz.steinmetz.UUCP> sunset!oconnor@steinmetz.UUCP writes: > ---------------- MICRO-processors wont hit 1000MIPS until > the year 2005 or so, if even then. How many microprocessor > architectures from 1965 are still around today? :-) > Easy one, answer is ZERO. 8-) The new word those days was "integrated curcuit". This was 23 years ago. Look 15 years forward and who knows if even bits count then. New ways to represent data are coming up all the time. Many of them can be used for storing data in more compact and/or different ways. I think that you must split this problem in two parts. One for them who needs more bits to represent larger numbers, and one fore the guys who wants to work with large addressing spaces. There is an analogy: New algorithms must be developed to increase computing speed, not only the raw processor power. New ways to represent data are requiered to decrease the amount of storage needed to represent it.