Quantization compresses FP32 weights to INT8 with big speed and memory wins, but a calibration set mismatched to your scene collapses accuracy. The calibration set is the soul of quantization Calibrate from real cameras: 200-500 frames covering day/night, weather and angles beat generic datasets; Per-layer sensitivity: keep accuracy-critical head/tail layers in FP16 mixed quantization; Always […]
Linux lands most interrupts on CPU0 by default, leaving one core saturated while others idle and total throughput capped. Interrupts defaulting to CPU0 is the trap Inspect: cat /proc/interrupts shows which core each IRQ lands on — watch NIC and storage controllers; Pin them: set /proc/irq//smp_affinity to spread NIC interrupts across the A76 cores; Pair […]
Zigbee2MQTT keeps dozens of sensors on local MQTT with no vendor cloud. On RK3588, the coordinator and channel plan decide reliability. Coordinator and channel Prefer ConBee II/III or SONOFF ZBDongle-E: solid Linux support and thorough community docs; Plan channels around Wi-Fi: Zigbee channels 15/20/25 avoid 2.4 GHz Wi-Fi’s 1/6/11 and cut dropouts sharply; Open permit_join […]
Changing the main SoC means redoing packages, power, DDR topology and thermals — that’s a respin, not a component swap. Budget it as a new project. An upgrade touches the whole board Is it worth it: if the bottleneck is NPU or multi-core CPU, yes; if it’s storage or networking, a new SoC changes nothing; […]
Subtitles are where most Jellyfin setups break: embedded tracks not showing, styled ASS turned into white boxes, burn-in causing stutter. Every one of these traces back to the transcode path. Which subtitle path to pick Embedded SRT/PGS passes through without transcoding — prefer muxing subtitles into the mkv container; Styled ASS needs libass rendering; when […]
Many teams pick a model and then discover it won’t keep up. The right order: throughput → per-frame budget → required compute → part number. Compute takt before compute power Reverse-engineer takt: at 60 parts/minute across 4 stations each inference gets roughly 200ms — reserve half for capture and post-processing; Cameras and buses: GigE Vision […]
The team’s pain points were third-party custody of files, slow large uploads, and per-seat pricing that scaled linearly. Why leave public cloud storage Solution: Nextcloud with OnlyOffice on dual-disk RAID1, with daily incremental backup to an external drive; Cost: hardware was a one-time spend roughly equal to two years of subscription, with zero monthly fees […]
WireGuard wins on minimal config and near-native speed. RK3588 as the server makes a phone in a café feel like it’s plugged into your LAN. Three steps Server: generate a keypair with wg genkey, open UDP 51820 and forward it; Clients: one keypair per device with AllowedIPs=10.0.0.0/24 so only LAN traffic tunnels; Remote sites: two […]
Most RKNN conversion failures trace to toolchain/driver version mismatch. Align versions first, then argue about the model. Version matching is the first gate Version triangle: PC rknn-toolkit2, on-device rknn_server/runtime and the kernel RKNPU driver must be mutually compatible; Operator check: read the verbose build output to confirm nothing falls back to CPU; Output alignment: RKNN […]
Default governors may not match your workload — sustained full load and bursty work want completely different policies. Understand the governors first Policy choice: performance suits sustained load but runs hot; ondemand or schedutil suits bursts; powersave fits always-idle devices; Big.LITTLE scheduling: RK3588 has 4x A76 plus 4x A55 — pin realtime work to A76 […]