Oh yes

Can someone tell me the major/minor device numbers for the /dev/rknpu device? It wasn't created in the image that I'm using. I'm trying to determine if the built-in npu has enough functionality to test a llm: # cat /sys/kernel/debug/rknpu/version RKNPU driver: v0.9.8 # I'm running the latest of the linux-image-vendor-rk35xx=25.5.1 v6.1.115 $ uname -a Linux cm3588-nas 6.1.115-vendor-rk35xx #1 SMP Wed Aug 13 03:55:09 UTC 2025 aarch64 GNU/Linux $ # LD_LIBRARY_PATH=/usr/lib/aarch64-linux-gnu python3 server.py Loaded RKLLM library from '/usr/lib/aarch64-linux-gnu/librkllmrt.so' RKLLM_LOG_LEVEL: 1 Initializing model with parameters: Path: /src/models/tinyllama-v1-rk3588/tinyllama-v1.0-1.1B-rk3588.rkllm Max Context: 30000 Max New Tokens: -1 N_Keep: 32 Use GPU: True Is Async (Library Flag): False E RKNN: [23:47:17.469] failed to open rknpu module, need to insmod rknpu dirver! E RKNN: [23:47:17.469] failed to open rknn device! E RKNN: [23:47:17.570] Device is not available Get device properties failed Error: rkllm_init failed (code 1). Failed to initialize RKLLM model. Server not starting. # ls -l /dev/rknpu ls: cannot access '/dev/rknpu': No such file or directory #

I looked a little closer and the included cpufrequtil init script is totally unaware of SoC clustering. If you put in a max value for the A76, it won't work for the A55. So I threw the problem into AI and asked for a flexible SoC solution (I'm of course simplifying it). This is what it gave me. Lightly tested: per cluster governors, per cluster min/max speed, ability to specify "min" and "max" rather than numeric values. cpufrequtils-default -> /etc/default/cpufrequtils cpufrequtils -> /etc/init.d/cpufrequtil YMMV. cpufrequtils cpufrequtils-default

You're right - this works perfectly. I'm not typically a Debian user, so I'm not sure where this should be configured in a proper way. But at least I can unleash the CPUs now. Thank you for the very helpful reply.

Hi all, I’m working with an RK3588-based board and running into a frequency scaling issue. All cores (cpu0–7) are stuck at 408 MHz, despite available OPPs going much higher (up to 2.35 GHz on the big cores). Attempts to raise frequencies via scaling_setspeed or userspace governor have no effect. The image I'm using is: https://www.armbian.com/nanopc-cm3588-nas/ (https://dl.armbian.com/cm3588-nas/Bookworm_vendor_minimal-omv) Device tree file is: fdtfile=rockchip/rk3588-friendlyelec-cm3588-nas.dtb # for cpu in /sys/devices/system/cpu/cpu[0-9]*; do echo "== $(basename $cpu) ==" cat $cpu/cpufreq/scaling_available_frequencies cat $cpu/cpufreq/cpuinfo_max_freq done == cpu0 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 1800000 == cpu1 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 1800000 == cpu2 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 1800000 == cpu3 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 1800000 == cpu4 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 2016000 2208000 2352000 2352000 == cpu5 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 2016000 2208000 2352000 2352000 == cpu6 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 2016000 2208000 2352000 2352000 == cpu7 == 408000 600000 816000 1008000 1200000 1416000 1608000 1800000 2016000 2208000 2352000 2352000 root@cm3588-nas:/# Getting DVFS/EM kernel message about inefficient OPP: [ 14.038585] cpu cpu0: EM: OPP:816000 is inefficient [ 14.048575] cpu cpu4: EM: OPP:1008000 is inefficient [ 14.048584] cpu cpu4: EM: OPP:816000 is inefficient [ 14.060709] cpu cpu6: EM: OPP:1008000 is inefficient [ 14.060718] cpu cpu6: EM: OPP:816000 is inefficient PVT readings [ 14.002363] cpu cpu0: pvtm=1425 [ 14.002523] cpu cpu0: pvtm-volt-sel=1 [ 14.011254] cpu cpu4: pvtm=1645 [ 14.015591] cpu cpu4: pvtm-volt-sel=3 [ 14.025824] cpu cpu6: pvtm=1665 [ 14.030065] cpu cpu6: pvtm-volt-sel=3 Attempting to raise speed: # for cpu in /sys/devices/system/cpu/cpu[0-9]*; do echo 1800000 | sudo tee $cpu/cpufreq/scaling_setspeed; cat $cpu/cpufreq/scaling_cur_freq; done 1800000 408000 1800000 408000 1800000 408000 1800000 408000 1800000 408000 1800000 408000 1800000 408000 1800000 408000 Why might all cores be capped at 408 MHz despite available OPPs and voltage steps?