Repository navigation
Replies: 2 comments 1 reply
|
This is almost certainly a power-delivery fault, not an ONNX Runtime bug — and the evidence in your report points at it more strongly than at the software. Here's why, and what would settle it. Why "5 V / 5 A was stable at 1 A" doesn't rule it out. Multithreaded inference doesn't raise average current much; it raises transient current. When several cores ramp simultaneously on an ARM SoC, the load step happens in microseconds and the supply sees a di/dt the VRM and cable inductance cannot follow. A bench meter (or a USB power meter) averages over milliseconds and will show a flat 5 V / 1 A while the rail is dipping below the SoC's minimum for tens of microseconds at a time. Allwinner boards with a single 5 V barrel/USB-C input are especially exposed because the SoC, DRAM and eMMC share one rail with little bulk capacitance. Complete power-off with no kernel log, no OOM, no thermal trip, and LEDs/fan going dark is the signature of a brownout: the PMIC's under-voltage lockout fires and cuts the SoC, so nothing gets a chance to log. A kernel panic, a thermal shutdown, or an OOM kill all leave evidence in Why multithreading correlates with it. More concurrent cores → larger simultaneous current step → deeper droop. That's exactly the pattern you describe: single-threaded Piper synthesis completes; The How to confirm it in one afternoon:
If it is the power path (most likely): add bulk capacitance at the 5 V input (a few hundred µF of low-ESR electrolytic plus ceramics), shorten and thicken the cable, and — most effective — check whether the board exposes separate power for the SoC vs. peripherals. If you're powering via USB-C from a laptop or a cheap hub, that alone can cause this; a dedicated 5 V supply with a 20 AWG cable often fixes it outright. If it isn't: the next discriminator is whether it reproduces under a pure CPU stress test ( One thing I'd flag for the ONNX Runtime maintainers reading this: the issue as filed has no ONNX Runtime-specific evidence, so it may reasonably be closed as a hardware/environment issue. If you want it kept open, include the |
|
Guyome41 — following up, because as filed this reads as a power-delivery fault rather than an ONNX Runtime bug: multithreaded inference raises transient current, not average current, and "stable at 1 A" doesn't test that. The stress-ng control and the bench-supply test in the comment separate them. If that matches what you found, please mark the comment above as the answer — it's the fast path to not spending weeks on the runtime. If you want the issue kept open against ONNX Runtime, include the stress-ng result; without it the maintainers can reasonably close it as hardware. |
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone,
I'm experiencing a serious issue with my Orange Pi Zero 3W (12 GB RAM, Allwinner A733, Debian Trixie, Linux kernel 6.6.98).
The board suddenly powers off completely during ONNX inference: all LEDs and the cooling fan turn off. There is no error message or useful kernel log, and I have to physically disconnect and reconnect the power supply to restart it.
Initially, I encountered the issue while running Piper TTS alongside Gemma using LiteRT-LM. However, I managed to reproduce the shutdown using ONNX Runtime alone, by directly running inference on the voice model, without Piper, audio playback, or Gemma.
What I've tested so far:
taskset), graph optimizations enabled/disabled, and CPU frequency limits.taskset. Explicitly settingintra_op_num_threads = 6avoids this behavior, but does not eliminate the shutdowns.The failures are intermittent, and some configurations complete several inference runs before the board suddenly shuts down.
Since this can happen with ONNX Runtime alone, it does not appear to be specifically caused by Piper, the I²S audio driver, or Gemma.
I'm wondering whether this could be related to CPU scheduling, power management, a kernel/driver issue, or SoC stability under certain workloads.
Has anyone experienced similar complete power-offs on the Orange Pi Zero 3W / Allwinner A733, particularly when using ONNX Runtime or multithreaded CPU inference?
Any suggestions would be greatly appreciated. Thanks!
All reactions