New boards: first runs (2026-11-11)

First bench and stress runs on the three Nucleo boards from the DigiKey order. Same firmware, libraries (libs.lock), toolchain image and adapters as on the nRF5340; only the board layer (firmware/common/<board>/) differs. Records are in ~/lab-runs/ on lib-lab1.

Boards

Board Core Clock Cycle counter Probe Port (power cycling)
NUCLEO-G0B1RE Cortex-M0+ 64 MHz (HSI16 + PLL) SysTick (no DWT on M0+) ST-LINK/V2-1 monitor 1, hub 4B0708619185 port 2 (yes)
NUCLEO-G474RE Cortex-M4F 170 MHz (HSI16 + PLL, boost) DWT CYCCNT STLINK-V3 monitor 1, port 3 (yes)
NUCLEO-H723ZG Cortex-M7 64 MHz (HSI, reset default) DWT CYCCNT STLINK-V3 monitor 2, hub 9E0C0069B5FA port 2 (yes)

The H723ZG runs at its reset clock on purpose: 550 MHz needs a supply configuration that matches the board, and a wrong one can lock the chip. Cycle counts are what the bench reports, so the comparison holds; flash numbers with the cache off include the reset-default wait states.

Bench (flash, cache/accelerator on, gcc -O2, newlib-nano memcpy)

Mean cycles, 1-byte push / pop · 64-byte block, cycles per byte. All 25 adapters ran with 0 FIFO errors on every board and code location (flash, SRAM, and CCM on the G474 / ITCM on the H723).

Library NUCLEO-G0B1RE (M0+, 64 MHz) NUCLEO-G474RE (M4, 170 MHz) NUCLEO-H723ZG (M7, 64 MHz HSI)
pxq RingBuffer_PackedState (C++) 35 / 38 · 10.2 20 / 21 · 7.5 17 / 16 · 3.5
lwrb (C) 263 / 256 · 13.5 180 / 176 · 5.1 145 / 136 · 3.8
AndersKaloer Ring-Buffer (C) 73 / 50 · 27.3 57 / 41 · 20.1 33 / 26 · 9.2
jnk0le Ringbuffer (C++) 36 / 42 · 11.8 23 / 26 · 9.5 35 / 35 · 6.6
ETL queue_spsc_isr (C++) 73 / 68 · 53.6 40 / 45 · 31.9 42 / 45 · 28.5
ETL queue_spsc_atomic (C++) 54 / 51 · 40.0 30 / 27 · 23.9 38 / 37 · 26.2
QuantumLeaps LFRB (C) 69 / 61 · 79.9 46 / 46 · 53.4 46 / 41 · 35.7
EmbeddedArtistry cbuf threadsafe (C) 74 / 60 · 81.9 56 / 46 · 58.4 37 / 31 · 25.4
ece-mohammad RingBuffer (C) 111 / 109 · 12.4 81 / 80 · 4.2 51 / 50 · 3.0
micoring SPSC mode (C) 127 / 131 · 128.2 98 / 97 · 96.9 72 / 71 · 59.9
kostyk348 lf_spsc (C) 159 / 154 · 212.6 72 / 67 · 79.4 65 / 59 · 56.2
diffstorm cb (C) 122 / 131 · 141.4 72 / 69 · 77.9 51 / 63 · 48.4
RIOT tsrb (C) 65 / 64 · 30.1 52 / 51 · 19.6 46 / 44 · 11.4
Contiki-NG ringbuf (C) 61 / 53 · 71.9 45 / 42 · 51.9 31 / 29 · 22.2
rigtorp SPSCQueue (C++) 52 / 83 · 73.2 35 / 49 · 45.5 36 / 43 · 29.1
wizard97 as shipped (C++) 41 / 41 · 55.9 29 / 25 · 33.4 24 / 22 · 16.5
wizard97 + PRIMASK macros (C++) 57 / 61 · 73.9 40 / 32 · 43.9 52 / 49 · 43.9
polycratia spscring (C++) 39 / 46 · 11.8 24 / 28 · 9.5 38 / 32 · 5.2
rt-spsc (C++) 45 / 44 · 59.4 26 / 24 · 31.9 36 / 36 · 28.9
TinyUSB tu_fifo (C) 138 / 131 · 12.3 101 / 100 · 4.2 57 / 60 · 3.5
pxq Topology::None, single-context (C++) 35 / 38 · 10.2 19 / 20 · 7.5 23 / 21 · 3.7
micoring default, single-context (C) 103 / 98 · 90.9 78 / 75 · 66.5 49 / 53 · 29.0
EmbeddedArtistry cbuf basic, single-context (C) 87 / 80 · 98.4 40 / 36 · 43.9 33 / 29 · 24.9
ETL etl::queue, single-context (C++) 46 / 47 · 61.4 32 / 34 · 38.9 25 / 26 · 19.4

Not yet checked: the read overheads BOARD_READ_OVERHEAD (G0B1 2, G474 2, H723 1) are assumed from the core manuals; the empty-call calibration in the records (raw minimum 9 / 8 / 2 cycles in flash) is consistent with them, but an independent check against a known instruction sequence (as nrf5340-cyccnt did for the M33) is still to do. Until then these numbers are preliminary.

Stress (ISR <-> main, 50 000 values per library and direction)

~/lab-runs/stm32-stress-first-20261111T191247Z: single, block and zero-copy (dma) on each board, 9 runs, all valid: real libraries error-free, the negative control caught in every run, nothing starved. wizard97 as shipped (not ISR-safe by its own docs) fails on all three boards (761–7989 errors), as on the nRF5340. Interrupt every 160–668 core cycles (prescaler 0).

These are short runs, not a campaign.

Fix found on the way

The first G474 run stalled: TIM2 is a 32-bit timer, and when the interrupt acknowledged a new period that the counter had already passed, it counted on to 2^32 (about 100 s at 170 MHz / 4) before the next interrupt. common/stm32/tim2_stress.h now restarts the counter in that case (commit 6a58fbc), which matches the nRF5340's 8-bit timer behaviour. Two records made while an old test process still used the same board were discarded (~/lab-runs/invalid-overlap-20261111/).

Multi-producer / multi-consumer stress (evening)

A second timer (TIM3, common/stm32/tim3_stress_b.h; base, enable bit and IRQ from ST's cmsis-device-g0 7c77b7cb, -g4 626ee412, -h7 81db1ec6) lets rb-mp-stress run on the three boards. 200 000 values per library, TIMER_PRESCALER 1 (~/lab-runs/stm32-mp-first-20261111T194458Z):

Board MPSC SPMC
G0B1RE (M0+) 0 errors, control caught 0 errors, control caught
G474RE (M4) 0 errors, control caught 0 errors, control caught
H723ZG (M7) 0 errors, control caught 0 errors, control caught

0 errors = pxq (MPSC/SPMC and MPMC), RIOT tsrb, wizard97 + PRIMASK macros.

The first G0B1 MPSC attempt failed because its run.sh was edited while it ran; it was rerun cleanly.

SAM D20 Xplained Pro (not working yet)

Board layer written (common/samd20/): 8 MHz OSC8M, 0 wait states, SysTick timing, TC0/TC2 as the two stress timers, flash image only (32 KB RAM). Register values from Zephyr hal_atmel 8f95c79f (Microchip's headers). probe-rs cannot drive the board, so common/samd20/bin/probe-rs stands in with OpenOCD. Bench, stress and MP firmware build. On the board, SWD fails with "stalled AP operation" (also before our firmware was ever flashed), and a USB power cycle of its port did not clear it; it worked once after a manual replug. adapter serial crashes OpenOCD 0.12.0 with this EDBG, so the stand-in does not select the probe by serial. To check with the board in hand.

Not done yet