About the other tools. OpenOCD and probe-rs are open-source projects that support far more chips and probes than Coil, and the ESP32 community relies on both. Coil does one thing, ESP32 over USB-JTAG and the ESP-Prog, which is part of why it can be faster there. If you work on either tool and think we set it up wrong, open an issue and we'll rerun it.
ESP32-C3 RISC-V · built-in USB-JTAG
2026-11-19. ESP32-C3 board running a small FreeRTOS test app; memory tests in RAM at 0x3FCA0000. OpenOCD v0.12.0-esp32-20260424 (ESP-IDF's build) at 40 MHz, probe-rs 0.32.0 asked for 40 MHz, Coil 0.2.0.
| Operation | Coil | OpenOCD | probe-rs | Coil vs the faster other |
|---|---|---|---|---|
| Attach | 6.47 msp95 6.83 ms | 19.7 msp95 27.2 ms | 28.1 msone run | 3.0× faster |
| Step | 2.83 msp95 2.99 ms | 15.6 msp95 17.4 ms | 23.2 msp95 27.5 ms | 5.5× faster |
| Step + registers | 2.88 msp95 3.11 ms | 26.3 msp95 35.0 ms | 52.1 msp95 59.1 ms | 9.1× faster |
| Registers after a step | 0.06 msp95 0.18 ms | 12.2 msp95 14.6 ms | 28.8 msp95 32.8 ms | 217× faster |
| Halt a running core | 4.55 msp95 5.22 ms | 16.1 msp95 23.8 ms | 14.7 msp95 22.4 ms | 3.2× faster |
| Memory read, 16 KiB | 117 KiB/s137 ms | 22 KiB/s732 ms | 24 KiB/s660 ms | 4.8× faster |
| Memory write, 16 KiB | 117 KiB/s137 ms | 126 KiB/s127 ms | 11 KiB/s1.4 s | 1.1× slower |
OpenOCD writes memory a little faster than Coil on the C3 (both are close to what the chip's USB link carries); probe-rs's attach was timed once, because its GDB server ends with the first client.
ESP32-C6 RISC-V · built-in USB-JTAG
2026-11-06/07. ESP32-C6 dev kit; memory tests in RAM at 0x40840000. The C6's USB-JTAG runs at 24 MHz at most, and both OpenOCD v0.12.0-esp32-20260424 and probe-rs 0.32.0 ran at 24 MHz.
| Operation | Coil | OpenOCD | probe-rs | Coil vs the faster other |
|---|---|---|---|---|
| Attach | 4.98 msp95 5.72 ms | 124 msp95 384 ms | 39.3 msone run | 7.9× faster |
| Step | 2.06 msp95 2.22 ms | 75.8 msp95 97.6 ms | 70.5 msp95 89.4 ms | 34× faster |
| Step + registers | 2.11 msp95 2.23 ms | 132 msp95 150 ms | 167 msp95 197 ms | 63× faster |
| Registers after a step | 0.10 msp95 0.19 ms | 48.7 msp95 62.1 ms | 94.5 msp95 128 ms | 483× faster |
| Halt a running core | 4.18 msp95 4.58 ms | 76.2 msp95 90.6 ms | 21.8 msp95 36.4 ms | 5.2× faster |
| Memory read, 16 KiB | 144 KiB/s111 ms | 3.5 KiB/s4.6 s | 5.6 KiB/s2.9 s | 26× faster |
| Memory write, 16 KiB | 143 KiB/s112 ms | 84 KiB/s191 ms | 3.2 KiB/s5.1 s | 1.7× faster |
OpenOCD and probe-rs were measured with the board's earlier firmware, Coil with the small test app it runs now; with the earlier firmware, Coil's step + registers was 3.4 ms (OpenOCD 132 ms). probe-rs's attach was timed once.
ESP32-S3 Xtensa, two cores · built-in USB-JTAG
2026-11-06/07. An ESP32-S3 board running a larger application; memory tests in RAM at 0x3FC88000; 40 MHz. OpenOCD v0.12.0-esp32-20260424, probe-rs 0.32.0.
| Operation | Coil | OpenOCD | probe-rs | Coil vs the faster other |
|---|---|---|---|---|
| Attach | 4.03 msp95 4.67 ms | 35.0 msp95 49.0 ms | 94.9 msone run | 8.7× faster |
| Step | 16.4 msp95 17.0 ms | 237 msp95 262 ms | 30.5 msp95 43.6 ms | 1.9× faster |
| Step + registers | 16.7 msp95 17.2 ms | 233 msp95 246 ms | 129 msp95 157 ms | 7.7× faster |
| Registers after a step | 9.72 msp95 9.95 ms | 15.9 msp95 29.9 ms | 94.9 msp95 111 ms | 1.6× faster |
| Halt a running core | 2.77 msp95 3.26 ms | 230 msp95 235 ms | not measured | 83× faster |
| Memory read, 16 KiB | 83 KiB/s194 ms | 49 KiB/s326 ms | 3.7 KiB/s4.3 s | 1.7× faster |
| Memory write, 16 KiB | not measured | 73 KiB/s218 ms | 7.8 KiB/s2.1 s | – |
OpenOCD's halt ran 8 times instead of 20 (the firmware stopped by itself after that). probe-rs stopped answering while the harness restored RAM after its write test, so its halt wasn't reached; its attach was timed once. Coil's memory write wasn't measured on the S3.
ESP32 (classic) Xtensa, two cores · ESP-Prog (FTDI)
2026-11-09. ESP32-MINI-1 board through an ESP-Prog. Coil on FTDI's own Windows driver at 15 MHz; OpenOCD v0.12.0-esp32-20260304 at 20 MHz, its fastest setting here (it needs the WinUSB driver for the ESP-Prog). probe-rs wasn't measured through the ESP-Prog.
| Operation | Coil | OpenOCD | probe-rs | Coil vs the faster other |
|---|---|---|---|---|
| Attach | 2.50 msp95 3.03 ms | 3.30 msp95 5.98 ms | not measured | 1.3× faster |
| Step | 3.40 msp95 3.83 ms | 16.0 msp95 29.3 ms | not measured | 4.7× faster |
| Step + registers | 3.41 msp95 4.15 ms | 16.5 msp95 31.0 ms | not measured | 4.9× faster |
| Registers after a step | 1.68 msp95 1.97 ms | 2.79 msp95 4.26 ms | not measured | 1.7× faster |
| Halt a running core | 1.36 msp95 1.54 ms | 17.7 msp95 20.7 ms | not measured | 13× faster |
| Memory read, 16 KiB | 500 KiB/s32.0 ms | 442 KiB/s36.2 ms | not measured | 1.1× faster |
Through the ESP-Prog, OpenOCD is close to Coil on attach and memory reads; the difference is in stepping and halting.
Flashing
Writing an ESP-IDF build to flash and checking it. Coil writes over JTAG through the chip's ROM, compressed, and skips every 64 KB chunk the flash already holds; esptool is ESP-IDF's own flasher, over the chip's USB serial port.
ESP32-C3: test app, bootloader + partition table + app (166 KB)
| Case | Coil | OpenOCD | probe-rs | esptool |
|---|---|---|---|---|
| Full write and verify | 1.72 s | 3.52 s | 2.07 s | 1.84 s |
| Nothing changed | 0.44 s | 0.63 s | not measured | 1.84 s (always rewrites) |
2026-11-19, five runs each, medians. Coil over JTAG (one of its five full writes took 6.7 s); OpenOCD v0.12.0-esp32-20260424 program_esp … verify for the three regions at 40 MHz; probe-rs 0.32.0 download of the app's ELF (it writes its own bootloader and partition table); esptool v5.3.1 from ESP-IDF v6.0.2 over the C3's USB serial port at 460800 baud. On the C3, esptool is about as fast as Coil for a full write.
ESP32-C6: test app (156 KB app)
| Case | Coil | idf.py flash | esptool |
|---|---|---|---|
| App edited | 1.3 s | 4.6–6.5 s | 1.8–2.2 s (app only) |
| Bootloader and app changed | 1.3–1.5 s | 4.6–6.5 s | 2.3–2.6 s |
| Nothing changed | 0.16–0.19 s | 4.6–6.5 s | 1.8–2.2 s |
2026-11-08, ESP-IDF v6.0.2, three runs each. idf.py flash includes its build check.
ESP32-S3: a 1.4 MB app
| Case | Coil | OpenOCD | esptool |
|---|---|---|---|
| Full write | 11.0–11.3 s | 30.7 s | 12.3 s |
| Small edit | about 3–4 s | 23.9 s | 12.3 s |
| Nothing changed | 2.6 s | 2.8 s | 12.3 s |
2026-11-06/08. OpenOCD program_esp v0.12.0-esp32-20260424 at 40 MHz; esptool over USB serial at 460800 baud. Coil's small-edit time is estimated from its per-chunk cost. OpenOCD skips an image only when all of it is unchanged.
ESP32 (classic) through the ESP-Prog: a 1.3 MB build, four regions
| Case | Coil | OpenOCD | esptool |
|---|---|---|---|
| Full write and verify | 10.7 s | 13.8 s | 55.7 s (serial, ROM routines) |
| Nothing changed | 0.5 s | 2.7 s | – |
2026-11-09. Coil on FTDI's driver at 10 MHz; OpenOCD v0.12.0-esp32-20260304 at 10 MHz, program_esp … compress, four regions. OpenOCD writes the app itself faster than Coil (4.7 s against about 9 s): it raises the chip's clocks first, which Coil doesn't do yet; Coil's total is lower because OpenOCD starts over for each region. esptool's own flasher stub can't flash this board, so it ran with the ROM's routines.
How we measure
- The harness speaks GDB's remote protocol directly to each debug server (Coil's
coil-gdb, OpenOCD, probe-rs'sprobe-rs gdb), so no IDE or GDB adds its own time. Each operation runs once to warm up, then 20 times; we report the median, and the 95th percentile underneath. - Same conditions: the same board, firmware, RAM address and USB port for all three, one tool at a time, Windows 10 on the same PC. Each tool picks the board by its serial number. Memory written by the test is put back and checked afterwards.
- Operations: attach (connect, negotiate, first stop); a single step; a step plus reading every register, which is what an IDE needs before it can show where you are; a register read on its own; halting a running core; reading and writing 16 KiB of RAM.
- Flashing times the whole command, from start to the board running again, 3–5 runs.
- Versions: each section names the versions and the JTAG clock each tool actually ran at. Results are rerun when a chip is added; older sections keep the versions they were measured with.