Combining Drum and Core
Hi, Brian! Yes, but the 96-bit floating point format was range-defined; being used, e.g., for DOUBLE PRECISION variables in...
John Savard
Sort of. Your average x86 CPU probably has around 20-120 instructions in flight at any given point in time. That would be at least 40-240 inputs, and 20-120 outputs. The average IPC may be in the neighborhood of 1 though, which would be 2 inputs and one output.
Yes, CPUs tend to only touch small amounts of memory at once. Of course, even if they could touch more, do you really want enough pins for all the accesses? Pins and pads are expensive.
John Savard) wrote, in part: Of course, I was thinking of the famous post by Robert Firth from comp.arch: ** Indeed it was. Here is one list, from the KDF9 programming manual...
Hrmm...caches perhaps?
I am not convinced that's true. What is the branch density on a given workload? I don't have my data with me, but ISTR observing anywhere from 5-25% branch density. Obviously, NT branches will continue to fetch linearly, but most loop branches will be taken backwards. I don't know about for commercial workloads.
So the real question is, how often are branches taken? I'm not sure of that answer, but perhaps you have data on this (or someone else does). I wouldn't just buttume that most branches are NT.
Moreover, I don't think it's realistic to ask programmers to actually try and deal with different types of memory with different access characteristics. I think this might work if the OS could do this transparently (with the help of the compiler-JIT). Programmers have enough problems trying to get threading right, I think it's pressing your luck to hope for another change. Although I agree that this should result in higher performance if you have different memory types for different access patterns. The real question is:
1. Can it be done in a way that is transparent to programmers? 2. If not, can-will programmers actually use it? 3. Is it worth the complexity and cost? What's the pay off? 4. Will such tricks narrow the general-purposeness of modern computers in an unacceptable manner?
Sure, there are a lot of techniques which produce higher performance for certain access patterns (i.e. to non-conflicting banks), which are easy to implement.
On 19-6-06 05:22, in article I still have my 40-year old KDF9 Programming Manual: "(C) All rights reserved English Electric Computers Limited price: five shillings" I can confirm that the manual says...
I'm not really familiar enough to comment on older architectures like the CDC6600.
DK
Alt Folklore Computers from Newsgroups