220 likes | 322 Views
Energy Optimization of Probabilistic Circuit. Yung-Chun Hu & Ching -Yi Huang 2013/08/05. Outline. Introduction PCMOS PBL Correctness analysis Probabilistic gate assignment PBC optimization Redundant removal Application Future work. Introduction. Without noise. Add noise.
E N D
Energy Optimization of Probabilistic Circuit Yung-Chun Hu & Ching-Yi Huang 2013/08/05
Outline • Introduction • PCMOS • PBL • Correctness analysis • Probabilistic gate assignment • PBC optimization • Redundant removal • Application • Future work
Introduction Without noise Add noise Raise voltage to avoid noise
Introduction Energy per switching: Energy ratio p
Introduction • Probabilistic OR: ∨p ,AND: ∧p , NOT: ¬p • Example: ∨p , assume p=0.9 A 0.9 F B
Correctness analysis probability RPG a Y 0.8 b 0.7 X c Z 0.9 ‧ ‧ ‧ 0.9 sampling result correct value simulation result super pattern input pattern -sample gate-sample
Testability • A Statistic-based Approach to Testability Analysis – (ISQED'08) Chuang-Chi Chiou • A Fast testability computation approach for each signal • Parallel Monte Carlo simulation
Testability 1 1 A 1 X 1 1 B Y 1 F Z 1 C D Testable
Testability 1 1 A 1 X 1 1/0 B Y 1 F Z 1 C D Testable
Testability 1 1 A 1 X 1 1/0 B Y 1 F Z 1/0 C D Testable
Testability 1 1 A 1/0 X 1 1/0 B Y 1 F Z 1 C D Testable
Testability 1 1 A 1 X 1 1 B Y 1/0 F Z 1 C D Testable
Testability 1 1/0 A 1 X 1 1 B Y 1 F Z 1 C D Testable
P-gate assignment Random assignment
P-gate assignment Assign probabilistic gates to those gates with lower testability
Redundant Removal • In deterministic circuit, we can inject SA faults and remove redundant sub-circuits. SA fault
Redundant Removal • In PBC, we may regard signals which have high probability to be 0 or 1 as stuck-at-faults. E A w 0.97 B C D If every PI has probability of 0.5 to be 1, w has probability of 0.97 to be 0.
Redundant Removal • In PBC, we may regard signals which have high probability to be 0 or 1 as stuck-at-faults. 0 Correctness=0.977 The node reduction is 66%.
Application • We use Matlab to model a decoder. • Generally, a grey value is a 8-bit value. • We assign probability for each pixel to toggle its value.
Application Assign weight to different Pos.
Future work • Energy analysis including transient and dynamic • Implement redundant removal