The Power Wall

Notes

• This is section 1.7
• For years, we could count on the clock speed increasing annually or semiannually
• The graph on page 40 of the book, however shows that this growth stopped about 2004
  • NOTE IN BOTH GRAPHS the y axis is logarithmic.
  • Check out y=x with logarithmic y scale on desmos.com/calculator
  • This graph shows power consumption decreasing about that point as well.
• But like most graphs in the book, it ends about 2012
• I found one on wikimedia commons: https://commons.wikimedia.org/wiki/File:CPU_clock_speed_and_Core_count_Graph.png
  • Not verified, but I believe it from observation
  • Note, the speed is still increasing but WAY SLOWER
  • But also note the core count starts rising as the clock speed delta decreases.
• Why?
  • We have reached our ability to cool the processor in the face of clock speed increase
  • AND as we are discovering, we are having problems? generating/providing sufficient power to support the increase
• power ∝ 1/2 × capacitive load × voltage² × frequency switched
  • Power and temperature are related.
  • Part of this equation discusses the power needed by a transistor
    • Energy ∝ Capacitive load × Voltage²
    • Capacitive load is the switch from 0 to 1 or 1 to 0
    • So switching costs energy
  • And the main equation says that if we switch faster, we need more power
• Over the years we have worked at reducing the power
  • And we did quite well for a while.
  • But our current transistor technology "leaks" power
  • Or it uses more power then just what is required to switch.
  • H&P state that 40% of power is leaked today (2017 version of the book, so at least 10 years old)
• I found a presentation on youtube I believe, "Why CPU Clock Speeds Stopped Increasing | The Power Wall & Dennard Scaling Explained", by Abhijit Pethe, at https://youtu.be/LH3Tvfx4Xbo?si=Kit0op78w88Wracf
  • I can verify the first part of this (pipelining, ...)
  • The second part matches H&P
• We still continue to work on this but we have switched to multicore
  • We use the additional transistors
  • But we don't switch any faster
• We can be faster only if we can use more cores
  • Naively, this works fine (a video playing, our game running and reading mail) where each task uses a core
    • Each task does not get faster
    • But we can do more tasks simultaneously at the same speed.
• But parallel programming is hard
  • Reasonable/Efficient parallel code is hard to write
    • Think of group work, generally one person spends quite a bit of time making the group function efficiently
  • Parallelism introduces several problems including non-determinism due to race conditions
    • Solutions to these are possible but they introduce more complexity
  • Parallelism also stresses other systems (memory, I/O, ...)
  • Finally, some portions of computation are inherently sequential
    • And that will eventually dominate in the computation
    • Amdahl's law.
• Customized hardware is partially the solution
• Other materials for semiconductor construction
  • Gallium Nitride, Silicone Carbide, Synthetic Diamonds
• And Quantum Computing, ...