@tengomucho @kmaincent @hcodina @avaliente-bc @rchevalier-brainchip @ktsiknos-brainchip
Hi, I am using an AKD1000 PCIe accelerator with a Raspberry Pi 5 running Ubuntu Desktop ARM64.
Environment:
- Host board: Raspberry Pi 5
- OS: Ubuntu Desktop ARM64
- Python: Python 3.12
- Python environment: venv
- AKD1000 driver repository: https://github.com/Brainchip-Inc/akida_dw_edma
- Driver branch: master
- Driver module: akida_pcie
- Device node: /dev/akida0
- Python package: akida runtime installed via pip
- Hardware: BrainChip AKD1000 PCIe accelerator
- lspci output includes:
BrainChip Inc AKD1000 NN coprocessor rev 01
Current status is below.
The AKD1000 device itself is detected and hardware inference works correctly.
I confirmed it is working
- lspci detects the AKD1000 PCIe device
- /dev/akida0 exists
- akida devices detects the device
- model.map(device, hw_only=True) succeeds
- Hardware-only inference runs successfully
- Inference statistics are printed normally
- Average framerate is reported, e.g. around 91.8 FPS
- Last inference clock is reported
- Last program clock is reported
The Problem is
Power measurement fails when enabling the SoC power monitor.
Code:
device.soc.power_measurement_enabled = True
Error:
RuntimeError: Unable to init INA: failed to send to bus: -2
After this error:
- device.soc.power_meter is None
- floor power cannot be read
- power event count is 0
- avg_power_mw remains null
- energy_per_frame remains null
I also observed that power_measurement_enabled may appear as true after setting it, but this does not seem reliable because the setter already failed with the INA initialization error and the actual power_meter object remains None.
Additional error observed when trying to print floor power:
TypeError: unsupported format string passed to NoneType.__format__
This seems to be a secondary error caused by floor power being None.
I already tried,
- Reinstalled the akida_dw_edma PCIe driver from the master branch.
- Removed old driver/module remnants before reinstalling.
- Confirmed that the active driver is the newer akida_pcie driver, not the old pedd_bc driver.
- Rebooted the system after driver installation.
- Checked that /dev/akida0 exists.
- Checked that the AKD1000 device is detected by akida devices.
- Checked for other processes using /dev/akida0.
- Confirmed that inference itself works in hardware-only mode.
- Confirmed that input dtype is not the issue by using uint8 input.
- Confirmed that inference duration is not the issue by running inference for several minutes.
- Confirmed that FPS and inference clock statistics are produced normally.
So the issue seems isolated to INA power monitor initialization, not AKD1000 inference itself.
--
Is INA-based power measurement supported on Raspberry Pi 5 + AKD1000 PCIe accelerator with the current akida_dw_edma master driver on Ubuntu ARM64?
If it is supported, are there any additional board-specific steps required, such as
- kernel configuration changes
- driver source modifications
- Raspberry Pi PCIe configuration
- I2C/bus configuration
- power monitor initialization settings
- specific akida runtime version requirements
Any guidance would be appreciated.
@tengomucho @kmaincent @hcodina @avaliente-bc @rchevalier-brainchip @ktsiknos-brainchip
Hi, I am using an AKD1000 PCIe accelerator with a Raspberry Pi 5 running Ubuntu Desktop ARM64.
Environment:
BrainChip Inc AKD1000 NN coprocessor rev 01
Current status is below.
The AKD1000 device itself is detected and hardware inference works correctly.
I confirmed it is working
The Problem is
Power measurement fails when enabling the SoC power monitor.
Code:
Error:
After this error:
I also observed that power_measurement_enabled may appear as true after setting it, but this does not seem reliable because the setter already failed with the INA initialization error and the actual power_meter object remains None.
Additional error observed when trying to print floor power:
This seems to be a secondary error caused by floor power being None.
I already tried,
So the issue seems isolated to INA power monitor initialization, not AKD1000 inference itself.
--
Is INA-based power measurement supported on Raspberry Pi 5 + AKD1000 PCIe accelerator with the current akida_dw_edma master driver on Ubuntu ARM64?
If it is supported, are there any additional board-specific steps required, such as
Any guidance would be appreciated.