Thermodynamic Perspective on Architectures

1
Thermodynamic Perspective on Architectures

• Sadasivan Shankar (Intel)
• Ack Ralph K. Cavin III (SRC), Victor V. Zhirnov
  (SRC)
• July 12
• Emerging Research on Architectures
• ERD Meeting, San Francisco

2
Main Points

• We have developed a general methodology for
  applying thermodynamic principles for information
  engines like the Carnot principle to heat engines
• More fundamental than simplistic capacitance
  based formalism currently being used
• The operating limit of 104 kT is possibly due to
  quantum state, architecture, and layout
• Two applications
• Similar to heat engines, will identify the ideal
  Compute Engine Carnots Compute Engine for
  ideal computing. This would serve as a limiting
  case for realistic architectures
• Estimate efficiencies for different architectures
  based on physics
• Layout efficiency
• Energy efficiency
• Potential trade-off between layout and energy
• Continue work on estimating minimum energy needed
  of various simple systems

3
Computing Engine Premise
Heat Engine
Computing Engine

• Similar to a heat engine, a computing engine can
  be visualized
• Goal is to use thermodynamics, which incorporates
  relations between systems components and
  determines the most energy efficient systems

4
A Perspective on Layout

• Ideal layout based on a single switch
• Interconnects and isolation may be driving in
  real systems

5
Entropy of a Single Switch System (1)

• Entropy is determined by statistical mechanics

Switching state
Quantum States
Nit Nit 2 is bit, Nit 4 is qit etc. kB -
Boltzmann constant N number of states in the
system (1 for single state switching) Et Total
Energy is estimated from statistical mechanics Z
Partition function
Switching Energy
Total energy
6
Free Energy of a Single Switch System

• Free energy is determined by thermodynamics

7
Simple Illustration

• For a binary switch, the minimum energy is
  determined by the need to maintain binary
  transition (bit) and energy of the particle in an
  isolated level
• For a classical switch, the following are the
  limits

Example Micro-Systems Bits N Pi EMin
Binary 1 D2/8a2 1/2N kT log 2
Inverter 1 D2/12a2 1/2N kT log 2
NOR 2 D2/24a2 1/4N 2kT log 2
NAND 2 D2/16a2 1/4N 2kT log 2
6-T SRAM 16 D2/48a2 1/216N 16kT log 2

• EMin is idealistic and is determined by the bits
  processed in the micro-systems

8
Ideal Architecture Comparison A Simple
Illustration
U 4 (14 x 10) Nit 16, 16, 16,16
U 8 (10 x 12) Nit 2, 2, 2, 2

• For an ideal classic switch, reducing minimum
  energy seems to be driven by U more than Nit (due
  to log Nit dependence)
• Indicates that fewer larger units with larger
  bits processed in each lower free energy for a
  classical switch
• Other factors not considered
• Interconnects, dissipation, functionality of
  architecture, unit layout, software, .