CS 7810 Lecture 15 A Case for ThermalAware

  • Slides: 14
Download presentation
CS 7810 Lecture 15 A Case for Thermal-Aware Floorplanning at the Microarchitectural Level K.

CS 7810 Lecture 15 A Case for Thermal-Aware Floorplanning at the Microarchitectural Level K. Sankaranarayanan, S. Velusamy, M. Stan, K. Skadron Journal of ILP, October 2005

Importance of Temperature • High power density cooling tech must improve Every additional watt

Importance of Temperature • High power density cooling tech must improve Every additional watt increases chip’s cooling/packaging cost by $4 • Higher temperature exponentially higher leakage • Temperature variations cause wear and tear

General Approaches • Reduce overall power consumption • Wait for some unit to reach

General Approaches • Reduce overall power consumption • Wait for some unit to reach temperature limit and then throttle back (dynamic thermal management) • Better floorplans so that heat is evenly distributed and likelihood of local hotspot is reduced Better floorplanning does not eliminate the need for DTM

Thermal Modeling • Thermal resistance: models the rate at which heat passes through •

Thermal Modeling • Thermal resistance: models the rate at which heat passes through • Thermal capacity: models the temperature rise because of heat absorption Resistance a thickness / area Capacitance a thickness x area

Models for Alpha-like Processor

Models for Alpha-like Processor

Effect of Lateral Spreading Thermal resistance = 0 Thermal resistance = infinity

Effect of Lateral Spreading Thermal resistance = 0 Thermal resistance = infinity

Simulation Setup • Ambient temperature of 40 o C • Trigger threshold (when DTM

Simulation Setup • Ambient temperature of 40 o C • Trigger threshold (when DTM is invoked) 111. 8 o C • Emergency threshold is 115 o C • Alpha core is 6. 2 mm x 6. 2 mm – cross-core latency is 4. 21 and 7. 16 cyc (on different metal layers) • L 2 cache is wrapped around the core to form a square

Simulated Annealing • In each iteration, perturb the solution with a probability that equals

Simulated Annealing • In each iteration, perturb the solution with a probability that equals the difference between the “quality” of current and optimal solution • What is the metric for “quality”? ( A + l. W ) T W = S cij dij A is area – some layouts increase the amount of white space W is the performance penalty becoz of wire delays T is peak temperature for the layout d is the distance between units in cyc c reflects the performance impact of that distance

Inputs to Algorithm • Profile a subset of benchmarks to estimate power density of

Inputs to Algorithm • Profile a subset of benchmarks to estimate power density of each block • Profile the performance impact of each set of wire delays to determine weights

Wire Delay Penalties

Wire Delay Penalties

Optimal Floorplans Flp-basic each wire delay gets an equal weight Dead space 1. 14%

Optimal Floorplans Flp-basic each wire delay gets an equal weight Dead space 1. 14% 5. 24% Total wire length is longer here Flp-advanced wire delays are weighted

Impact on Peak Temperature Additional temp. reduction becoz of poorer IPC and lower leakage

Impact on Peak Temperature Additional temp. reduction becoz of poorer IPC and lower leakage

Comparison with DTM (DVS) DVS: Assumes 10 ms overhead for a change DVS-i: Assumes

Comparison with DTM (DVS) DVS: Assumes 10 ms overhead for a change DVS-i: Assumes no overhead for a change Emergency threshold

Title • Bullet

Title • Bullet