Float benchmark on Raspberry Pi Pico

Files

• float_bench.c: benchmark program for Raspberry Pi Pico
• float_bench_arduino/ directory: benchmark program for Arduino-based environments, such as ESP32, for comparison.

Methods

The benchmark tasks are:

1. Generate 4000 Gaussian random numbers using Box-Muller method.
2. Calculate mean of them.
3. Calculate variance of them.

Results

Board	Chip	Arch	Clock [MHz]	FPU	SDK	Generation [us]	Mean calculation [us]	Variance calculation [us]
Raspberry Pi Pico	RP2040	Cortex-M0+	1251	No	Pico SDK	49081	11560	17100
Raspberry Pi Pico	RP2040	Cortex-M0+	?2	No	Arduino mbed core	248237	27350	36410
Raspberry Pi Pico	RP2040	Cortex-M0+	133	No	Arduino-Pico (by Earle Philhower)	44557	10420	15168
ESP32-DevKitC	ESP32	Xtensa LX6	240	Yes	Arduino	11965	3217	3334
AtomS3	ESP32-S3	Xtensa LX7	240	Yes	Arduino	9369	1138	1278
M5Stamp C3	ESP32-C3	RV32IMC	160	No	Arduino	135386	13638	19240
Seeeduino Xiao	ATSAMD21G18A-MU	Cortex-M0+	48	No	Arduino	679416	83082	113193
Wio Terminal	ATSAMD51P19	Cortex-M4	120	Yes	Arduino	29235	836	936
Nucleo F411RE	STM32F411RE	Cortex-M4	100	Yes	STM32duino	16284	1014	1139
Arduino UNO R4 WiFi	RA4M1	Cortex-M4	48	Yes	Arduino	81498	2094	2344

Note: Above values are results of a single measurement. "Mean calculation" and "Variance calculation" does not intend the mean/variance of results. These are time spent to calculate mean and variance of 4000 Gaussian random numbers.

Discussion

Performance difference between Pi Pico SDK and Arduino Mbed Core

On Raspberry Pi Pico, Arduino version was slower than Pico SDK version despite the chip was the same. It seems because of the difference of software float routines. Current version of Arduino Mbed Core does not use the optimized float routines built in RP2040 (at the time of October 23, 2021). However, it seems to be solved soon. The situation is similar in November 2023.

Footnotes

1. Default clock frequency (see page 45 of Raspberry Pi Pico C SDK document). Max 133 MHz. ↩

2. See https://github.com/arduino/ArduinoCore-mbed/issues/221 ↩