Boards

Which instruments play on which board, and how much room each leaves. Every instrument was timed on one board of each chip family, and three families can’t play them at all; the table under the first one says why.

What plays where

Each cell is the most notes held, of 1, 6 and 10, that render inside the block, and in brackets the share of the block left over at that count. “over” means even one note takes longer than the block, with what one note costs. Voices is how many the instrument was built with, up to the pump’s 24, and Heap the memory it takes once built with them, about the same on every board, so a board needs PSRAM or SDRAM to hold more than one or two. An instrument with one voice holds one note by design, and poly5, with five, was timed with one note.

Instrument

Voices

Heap

ESP32-P4

STM32H747

ESP32-S3

ESP32

RP2350

acid303

1

181 KB

1 (65%)

1 (88%)

1 (33%)

1 (2%)

1 (38%)

bass_pedals

1

188 KB

1 (81%)

1 (82%)

1 (62%)

1 (25%)

1 (36%)

clavichord

10

571 KB

1 (68%)

1 (66%)

over (100%)

over (179%)

over (131%)

combo_organ

24

1535 KB

10 (27%)

10 (31%)

6 (19%)

1 (2%)

1 (15%)

compact_organ

24

913 KB

6 (6%)

1 (49%)

1 (10%)

over (178%)

over (180%)

duo20

8

608 KB

1 (64%)

1 (72%)

1 (16%)

over (138%)

over (113%)

duosynth

1

305 KB

1 (60%)

1 (62%)

1 (17%)

over (159%)

over (154%)

electric_grand70

10

590 KB

6 (33%)

6 (23%)

over (115%)

over (186%)

over (105%)

electric_piano

10

546 KB

10 (53%)

10 (50%)

1 (53%)

1 (5%)

1 (19%)

electronic_drums5

8

588 KB

6 (63%)

6 (61%)

6 (19%)

over (109%)

1 (10%)

hybrid_mono

8

468 KB

1 (73%)

6 (33%)

1 (45%)

1 (11%)

1 (31%)

mono101

1

280 KB

1 (78%)

1 (79%)

1 (57%)

1 (17%)

1 (27%)

mono_one

8

584 KB

1 (68%)

1 (68%)

1 (28%)

over (135%)

over (123%)

mono_synth

1

289 KB

1 (74%)

1 (75%)

1 (46%)

1 (2%)

1 (11%)

patch2600

8

1478 KB

1 (66%)

1 (67%)

1 (27%)

out of memory

over (140%)

poly106

6

477 KB

6 (6%)

6 (10%)

1 (39%)

over (122%)

over (115%)

poly5

5

495 KB

1 (69%)

1 (70%)

1 (36%)

over (124%)

over (111%)

poly6

6

539 KB

1 (58%)

1 (59%)

1 (17%)

over (164%)

over (162%)

poly8

8

712 KB

1 (62%)

1 (60%)

1 (14%)

over (169%)

over (157%)

poly80

8

1008 KB

1 (4%)

1 (40%)

over (172%)

over (314%)

over (259%)

polya6

16

2119 KB

1 (37%)

1 (46%)

over (113%)

over (162%)

over (203%)

polyxa

8

649 KB

1 (66%)

1 (65%)

1 (27%)

over (144%)

over (135%)

reed_piano

10

667 KB

10 (8%)

10 (8%)

1 (57%)

1 (6%)

1 (14%)

reed_piano200a

10

646 KB

10 (57%)

10 (54%)

10 (3%)

1 (25%)

1 (31%)

rhythm606

7

626 KB

6 (64%)

6 (64%)

6 (17%)

1 (19%)

1 (33%)

rhythm78

15

734 KB

10 (48%)

10 (46%)

1 (49%)

over (113%)

over (102%)

rhythm808

16

862 KB

10 (11%)

10 (11%)

1 (50%)

over (113%)

1 (1%)

rhythm909

11

1043 KB

6 (28%)

6 (27%)

1 (47%)

over (118%)

over (105%)

state_variable_mono

8

840 KB

1 (69%)

6 (0%)

1 (32%)

over (130%)

over (113%)

string_ensemble

24

992 KB

1 (50%)

6 (6%)

over (102%)

over (206%)

over (211%)

tonewheel3

16

940 KB

6 (15%)

6 (20%)

1 (13%)

over (180%)

over (160%)

vocoder330

12

1468 KB

1 (48%)

1 (47%)

over (106%)

out of memory

over (223%)

west_coast_synth

8

810 KB

1 (74%)

6 (1%)

1 (46%)

over (106%)

1 (4%)

The boards, all on development builds of MicroPython 1.30 with audiodsp 0.6.16 compiled in, at 48 kHz in 128-frame blocks, so 2667 µs a block:

Column

Board

Clock

Memory

ESP32-P4

Waveshare ESP32-P4-WIFI6-DEV-KIT

400 MHz, two cores

32 MB PSRAM

STM32H747

Arduino GIGA R1 WiFi

480 MHz Cortex-M7

8 MB SDRAM

ESP32-S3

Waveshare ESP32-S3-Touch-LCD-7

240 MHz, two cores

8 MB PSRAM

ESP32

Adafruit QT Py ESP32 Pico

240 MHz, two cores

2 MB PSRAM

RP2350

Pimoroni Pico Plus 2 W

150 MHz Cortex-M33

8 MB PSRAM

On the ESP32 the instruments ran from .mpy files on the board’s file system rather than frozen into the firmware, so their code took heap as well, and with 2 MB of PSRAM polya6 built six voices, patch2600 had no room for its graph, and vocoder330 didn’t fit at all.

On the ESP32, S3 and P4 the instrument renders on its own core, so the space left over is spare. On the STM32H747 and the RP2350 the instrument shares the only core with your program, so the space left is all your program gets.

Chips that can’t play them

Chip

Board tried

Why not

RP2040

Waveshare RP2040-Touch-LCD-1.28

264 KB of RAM, of which about 224 KB is free once audiodsp is in, and no floating point unit. acid303 and bass_pedals fit and take 509% and 708% of the block for one note; every other instrument needs more heap than the chip has.

nRF52840

Seeed XIAO nRF52840 Sense

1 MB of flash: audiodsp with ulab is 382 KB larger than the space for firmware, before an instrument.

SAMD51

Adafruit NeoTrellis M4, on CircuitPython

CircuitPython has no audiograph, the node the instruments render on, and its 127 KB of free heap is under the 180 KB or so the smallest instrument takes.

How it was measured

performance.md has the method, and board-timings.csv every figure behind the table: per board, instrument and note count, the voices built, the heap taken, and the µs a block for the graph and for the whole step().