Montek Singh COMP Oct 11, 2011
Today’s topics: ◦ more on error metrics ◦ more applications ◦ architectures and design tools ◦ challenges and benefits ◦ open questions
For arithmetic units, error metric based upon an error threshold, δ ◦ errors < δ are tolerable ◦ p δ = prob (err < δ)
Applications which harness probabilistic behavior ◦ algorithms with repeated execution with the same inputs resulting in distinct outcomes (with some prob. distribution) Separate algorithm into deterministic and probabilistic parts
Bayesian Inference ◦ statistical inference technique mimicking human decision-making process ◦ set of hypotheses and probability weights ◦ each observation leads to a revision of prob weights ◦ Example:
Example ◦ Given prob of rain prob of sprinkler being on given rain ◦ Find: prob of rain given that the grass is wet Implemented using PCMOS
Random Neural Networks ◦ Poisson process models the “firing” of a neuron Probabilistic Cellular Automata ◦ Each cell’s next state is a function of its neighbors ◦ Next state could be 0 or 1 with certain prob Hyper Encryption ◦ Random seed generated using PCMOS
Applications that tolerate probabilistic behavior ◦ multimedia mostly ◦ signal processing ◦ others?
Different partitioning of deterministic vs. probabilistic parts of an algorithm
Host (deterministic) vs. coprocessor (probabilistic) partitioning
Comparison ◦ Host = deterministic ◦ without coprocessor: software only ◦ with coprocessor: using PCMOS using CMOS
Quality of randomness using PCMOS vs. pseudo RNG
Benefits ◦ …? Challenges ◦ …?
Architectural questions ◦ Which design is better in terms of E-p tradeoff? ◦ Example: Which adder is better: carry-skip or ripple-carry? carry-skip adder has faster propagation time ripple-carry adder consumes less energy But: carry-skip adder may be better when there are delay-induced errors! Design tools ◦ What type of tool support is needed? ◦ Simulation and validation?