CS 575 Supercomputing - Lecture Outline
Chapter 3: Memory

Sept. 22, 2003

Dr. Kris Stewart (stewart@sdsu.edu)
San Diego State University

This URL is stewart.sdsu.edu/cs575/lecs/ch3.html

Focus:
Speed of processors is increasing steadily, performance doubling every 18 months, agrees with Moore's Law (Time-dependent: speed of computers is increasing almost 10-times in less than five years). What Moore actually observed was that the density of transitors on a silicon Integrated Circuit (IC) doubles every 18 months, since 1962 (George Moore, Intel Engineer and CEO, retired).
Intel Research / Silicon / Moore's Law
Is Moore's Law Irrelevant? Electronic News 8/14/2003

So what does that really mean? (18 months = 1.5 Years)
Year 0 Year 1.5 Year 3 Year 4.5 Year 6
Power=1 Power=2 Power=4 Power=8 Power=16

  • Power is usually CPU processor speed, but could someday be
  • In terms of names for numbers:
    1. petabytes (PB) of storage (petabyte is 1015 bytes) for files and such, 
    2. gigabytes (GB) of memory (gigabyte is 109 bytes) for running programs.
      Note: Interpreter + user program + data *or* Compiler processes user program to produce object code which must all fit in memory; then object code + data runs in memory. More details in a couple of weeks when we cover Ch. 5 What a Compiler Does 
    3. teraflops (TFLOPS) - overall measure of processor speed (teraflop is 1012 flops, or Floating Point Operations per Second). More details \ next week when we cover Ch. 4 Floating Point Numbers 
    4. input-output (I/O) capacity of hundreds gigabits (Gb) per second.
      (NOTE: Gb and GB differ by nearly an order of magnitude (factor of 10)
    Memory performance doubling every seven years, therefore need alternate mechanisms for successful High Performance Computing, because memory speed increase just isn't keeping up with the processor speed increase.
    (p. 33 Text) Example Memory Technology Access Time: 1981 IBM XT 8088 access time of commodity DRAM (200ns) was shorter than the clock cycle time (4.77 MHz)

    4.77 MHz ? < ? 210ns


    What does that really mean? Let's convert 200 ns (nanoseconds) to MegaHertz since they are reciprocals:
    1/200 ns ~= 1/(200*10-9) = 1/2 * 1/10-7 = 0.5*10+7 = 5.0 E+6 this gives a clock cycle time that would allow roughly 5 Million Cycles per Second, i.e. 5MHz.
    What memory response is needed to keep up with 1.8 Ghz Pentium4?
    Let's work on the board. 1.8 GHz ~= 2 GHz ~= 1/?

    What memory response is needed to keep up with 330 MHz Pentium?
    1/[330*106] = 1/.33*109 =~ 1 / [1/3 * 109] = 3 ns memory chip -->
    Economics is a driving force in computing. Possible approaches to memory performance

    In this chapter we will focus on

    Compare VHS tape (rewinding and fast forward) to DVD.


    Fig. 3-1: Cache lines can come from different parts of memory

    © O'Reilly Publishers (Used with permission)

    Interleaved and Pipelined Memory Systems

    Fig. 3-9 Multibanked memory system

    © O'Reilly Publishers (Used with permission)

    Memory Stride Effect for Multibank Memory