1. Arani Bhattacharya, Han Chen, Peter Milder, Samir R. Das
Quantifying Energy and Latency Improvements of FPGA-Based Sensors for Low-Cost Spectrum Monitoring
Arani Bhattacharya, Han Chen, Peter Milder, Samir R. Das
Stony Brook University

2. Regulators Want to Monitor Spectrum Usage
Requires continuous monitoring over a large area at low cost
Multiple studies in Dyspan’ 17 and 18 and other conferences
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Crowdsourcing promises to satisfy these requirements

3. Practical Challenges of Crowdsourcing
Challenging to deploy outdoors
Cost (Energy)
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Energy constraint is a major hindrance to outdoor deployment

4. Spectrum Sensors Need to be Cheap
Requires low-cost sensor
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Requires low-cost processor with Software-defined Radio

5. A Representative Low-cost Processor
Raspberry Pi connected to RTL-SDR
$40
$20
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Raspberry Pi with RTL-SDR a common representative

6. Raspberry-Pi based sensors are widely used
Raspberry Pi connected to RTL-SDR
Used by multiple distributed spectrum sensing projects like SpecSense, Electrosense and RadioHound
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors

7. The Process of Sensing Raspberry Pi connected to RTL-SDR
Compute FFT on Raspberry Pi
Data goes to Raspberry Pi through USB Port
No need to discuss the current approach. Also change this spectrum or signal map.
SDR Senses Data
Crowdsourced sensors

8. Which Step Costs Most Energy?
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Compute step consumes 67% of total energy

9. Practical Challenges of Crowdsourcing
Low compute power
High Latency
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Latency is a major challenge

10. Which Step has High Latency?
Low compute power
High Latency
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors
Compute also has high latency

11. Inside a Spectrum Sensor
Raspberry Pi connected to RTL-SDR
No need to discuss the current approach. Also change this spectrum or signal map.
Crowdsourced sensors

12. How can we Reduce Energy and Latency?
Raspberry Pi connected to RTL-SDR
Same computation possible on hardware too!
No need to discuss the current approach. Also change this spectrum or signal map.

13. How can we Reduce Energy and Latency?
FPGA-board with SDR
Raspberry Pi connected to RTL-SDR
Replace by
No need to discuss the current approach. Also change this spectrum or signal map.
Replace Raspberry Pi with an FPGA board

14. FPGA performs computation in hardware
FPGA-board with SDR
Raspberry Pi connected to RTL-SDR
Replace by
No need to discuss the current approach. Also change this spectrum or signal map.
Computes in hardware
Hardware execution is 100 times faster and energy-efficient

15. Can Such Gains be Obtained in Practice?
FPGA-board with SDR
Android smartphone with RTL-SDR
Raspberry Pi connected to RTL-SDR
No need to discuss the current approach. Also change this spectrum or signal map.
We benchmark FPGA, smartphone and Raspberry Pi sensors

16. Benchmarking Latency FPGA-board with SDR
Android smartphone with RTL-SDR
Raspberry Pi connected to RTL-SDR
Computed by measuring the number of clock cycles
No need to discuss the current approach. Also change this spectrum or signal map.
Obtained using time command on actual computation

17. Latency Measurements
No need to discuss the current approach. Also change this spectrum or signal map.

18. FPGA-based sensor is at least 73 times faster
Impact on Latency
73 times faster
No need to discuss the current approach. Also change this spectrum or signal map.
FPGA-based sensor is at least 73 times faster

19. Measured using Monsoon Power Monitor
Benchmarking Power
FPGA-board with SDR
Android smartphone with RTL-SDR
Raspberry Pi connected to RTL-SDR
No need to discuss the current approach. Also change this spectrum or signal map.
Measured using Monsoon Power Monitor

20. Measured using Monsoon Power Monitor
Benchmarking Power
FPGA-board with SDR
Android smartphone with RTL-SDR
Raspberry Pi connected to RTL-SDR
Lots of unutilized components in general-purpose board
No need to discuss the current approach. Also change this spectrum or signal map.
Difficult to estimate!!!
Measured using Monsoon Power Monitor

21. Need an Accurate Way to Measure Power
FPGA-board with SDR
Android smartphone with RTL-SDR
Raspberry Pi connected to RTL-SDR
Simulator tool provided by manufacturer gives an accurate estimate
No need to discuss the current approach. Also change this spectrum or signal map.
Measured using Monsoon Power Monitor

22. Impact on Energy Consumption
No need to discuss the current approach. Also change this spectrum or signal map.

23. Impact on Energy Consumption
14 times less
No need to discuss the current approach. Also change this spectrum or signal map.
FGPA consumes 14 times less energy

24. Can Power Consumption be Further Reduced?
Simulate a smaller FPGA
Larger than needed
No need to discuss the current approach. Also change this spectrum or signal map.
Can a smaller board save more energy?

25. Impact on Energy Consumption
No need to discuss the current approach. Also change this spectrum or signal map.

26. Impact on Energy Consumption
29 times less
No need to discuss the current approach. Also change this spectrum or signal map.
Using a smaller board saves 50% more energy

27. Is Compute Still an Energy Bottleneck?
No need to discuss the current approach. Also change this spectrum or signal map.
Compute consumes around 12.5% of total energy

28. Is Compute Still a Latency Bottleneck?
No need to discuss the current approach. Also change this spectrum or signal map.

29. Is Compute Still a Latency Bottleneck?
No need to discuss the current approach. Also change this spectrum or signal map.
Compute now takes only 50% of total latency

30. Quantifying Energy and Latency Improvements of FPGA-Based Sensors for Low-Cost Spectrum Monitoring
Identified computation as bottleneck in spectrum sensors
Benchmarked FPGA-based sensors with Raspberry Pi and smartphone
Reduces latency by 73 times
Reduces energy consumption by 14 times
Showed that our FPGA-based sensor