Report

COMPARISON OF ADAPTIVE VOLTAGE/FREQUENCY SCALING AND ASYNCHRONOUS PROCESSOR ARCHITECTURES FOR NEURAL SPIKE SORTING May 2013 – EE241 Final Project Introduction to ‘Spikes’ ‘Spikes,’ also called ‘action potentials,’ are when the electrical potential of a neuron rises shortly. It is part of the communication protocol of the brain. Frequency Range 10-500Hz Duration 3-5ms Resolution 8-bit Spike Sorting Systems • Operate at constant 37°C • Extremely low power • Extremely low area Purpose Statement Adaptive Voltage and Frequency Scaling vs. Asynchronous Timing Determine the conditions • under which adaptive voltage and frequency scaling provides a better power performance than asynchronous timing in a neural spike processor. • under which asynchronous timing provides a better power performance than adaptive voltage and frequency scaling in a neural spike processor. Theory • E = I*VDD*t • If you can decrease voltage and/or current faster than the increase in delay for an operation, you save energy. • Asynchronous circuits process data at the same energy ‘every time.’ • When data rates are low, I and VDD may be reduced such that the energy per spike of a synchronous solution is lower than the asynchronous solution. • The issue is that at some point, leakage current dominates, making increased operating time undesirable. Verilog Function • Detection of the largest value in a spike and the point at which it occurs facilitates spike sorting. • Sorting spikes on chip results in a reduced amount of transmitted data. Synchronous Logic Prior Art Asynchronous Logic Data Collection Methodology Verilog Cadence hSpice sim LVS sim Synthesis *RC E vs. f Data sim .sp netlist Place and Route sim Logic Parasitic extraction Matlab Demonstrations of functionality with synch and asynch. Determination of max. frequency and E vs. VDD. Known Simplifications • Process Corners • Entire Verification Space • Noise/Margins • Critical Path Monitors • Adaptive control circuitry Spice Data: Energy/Op vs. Frequency Asynchronous: 300 less MOSFETS >5000 less Rs and Cs At 20kHz: Synch: 1.9E-9A – 7.6pJ Asynch: 1.71E-9A, 6.8pJ Spice Data: Current Breakdown - Synchronous Assumptions • Process Corners • Only used TT as effects should be linear on both architectures. • Verification Space • Test vector data was designed to be representative, not the worse case scenario. Testing across a wide range of frequencies and voltages is an open problem. • Noise/Margins • Assumed no noise for this analysis. Noise should hurt both circuits equally. Critical Path Monitor • Subthreshold critical path monitors are highly sensitive to PVT both globally and locally. • Current is exponential with changes in , , and T. • Implies a currently unachievable level of specificity. • First order approximation might be to double energy. Adaptive Control Circuitry, Clock, + Vdd •Assuming that the adaptive frequencies are less than the global clock, we can implement the change in frequency with a clock divider (counter). •Voltage scaling circuitry would take the global supply and reduce it using one of many possible techniques. –Linear Regulators: Inefficient for large changes in voltage. –Capacitive Regulator: Results in supply ripples. –Magnetic Regulator: Highest efficiency, but requires external parts Matlab: Transient State Machine at .5MHz Matlab: Transient State Machine at .6Mhz Matlab: Transient State Machine at 5Mhz Matlab: Total Energy + Vdd vs. Event Frequency Conclusion • Asynchronous logic operates at a higher frequency for all supply values than the synchronous version. Similarly, for an equal operating frequencies, asynchronous can operate at a lower VDD. • For the speed of a neural processor, both circuits operate at the lowest supply voltage and leakage current dominates. There is no means to perform adaptive analysis. • In an application where incoming data frequency is on the order of magnitude of the intrinsic asynchronous circuit operating frequency, the adaptive circuit requires less energy, as long as the average voltage required of the adaptive circuit is less than the intrinsic voltage of the asynchronous circuit. References and Acknowledgements Acknowledgements: Brian Zimmer, Nathan Narevsky, Jan Rabaey, Stevo Bailey, Tsung-Te Liu, Bora Nikolic APPENDIX Spice Data: Power vs. Max Operating Freq.