C0-microSD Benchmarks
The benchmarks below evaluate the speedup achieved by replacing Monte Carlo simulation of sensor calibration and conversion routines with UxHw-based computation running on the C0-microSD (Athens-16 representation), versus Monte Carlo executions on an ARM Cortex-M33 microcontroller (150 MHz, ~39mW).
The application kernels in the table correspond to the manufacturer-specified calibration algorithms that convert each measured ADC value into a calibrated sensor output.
The number of Monte Carlo iterations for each application is chosen such that the output probability distribution reaches the same degree of convergence as the UxHw Athens-16 result.
The following figures come from the Signaloid-Compute-Module-Demo-Sensor demo.
| Application | Monte Carlo iterations | Monte Carlo latency | UxHw latency | Speedup |
|---|---|---|---|---|
| Flusso FLS110 | 291k | 3 s | 1.63 s | 1.8× |
| NXP MPX4100A | 185k | 0.7 s | 0.42 s | 1.6× |
| NXP MPXx6250A | 215k | 0.8 s | 0.42 s | 1.9× |
| Sensirion SDP3x | 193k | 2.3 s | 0.42 s | 5.5× |
| Sensirion SFM3100 | 2.7M | 8.2 s | 23.56 ms | 350× |
| Sensirion SHT3xARP | 193k | 1.6 s | 1.26 s | 1.3× |
| Sensirion SHT4xI | 204k | 1.7 s | 1.26 s | 1.4× |
UxHw latency is the latency for reaching the same degree of convergence of the distributional result as the multi-iteration Monte Carlo.
Next steps
- UxHw in Silicon, why distributional execution replaces the Monte Carlo iterations counted above.
- Specifications, the compute, memory, and power figures behind these results.
- Examples and Demos, the demo applications these measurements come from.
- Getting Started, running your own workload on the module.