If you are building a private, subscription-free NVR and you have typed “best SBC with built-in NPU for Frigate” into a search box, you already know the Google Coral USB is no longer the easy answer it used to be. The Coral gasket driver was archived in April 2026, modern Linux kernels past 6.2 will not build it without community patches, and Frigate’s own hardware page quietly dropped it from the top recommendation. That left a clear gap, and a whole new class of single-board computers with neural accelerators soldered onto the SoC has stepped straight into it.
I have been running a Frigate build on a Radxa Rock 5B for nine months, and our team added five other boards to the bench over the last quarter. We measured inference speed on the same four Dahua IP cameras, watched how each board handled go2rtc re-streaming, and tracked idle watts with a Kill-A-Watt. This guide is the result: the six best SBCs with a built-in NPU for Frigate right now, ranked by what actually matters in a 24/7 home NVR: real-world inference time, RAM headroom, detector support, and power draw.
If you want a quick refresh on Frigate-compatible PoE cameras before you start, our 8 Best PoE Cameras for Frigate NVR guide pairs well with this list. For a closer look at budget USB accelerators as a fallback, see the 8 Best NPU Accelerators for Frigate on a Budget roundup.
Table of Contents
Top 3 Best SBCs with a Built-In NPU for Frigate in September
Best SBCs with a Built-In NPU for Frigate in 2026
| Product | Specs | Action |
|---|---|---|
 Radxa Rock 5B 8GB (RK3588) |
|
Check Latest Price |
 Radxa Rock 4D 6GB (RK3576) |
|
Check Latest Price |
 Vilros Raspberry Pi 5 AI Kit |
|
Check Latest Price |
 Orange Pi 5 Plus 8GB |
|
Check Latest Price |
 Raspberry Pi AI HAT+2 (Hailo-10H) |
|
Check Latest Price |
 Orange Pi 4 Pro 4GB |
|
Check Latest Price |
1. Radxa Rock 5 Model B (RK3588) – Best SBC With Built-In NPU for Frigate Overall
Official Rock 5 Model B 5B 8GB Single Board Computer Rockchip RK3588, 8Kp60, On/Off Button, eMMC Socket, 40-pin GPIO, Debian/Ubuntu
Rockchip RK3588 SoC
6 TOPS NPU
8GB LPDDR4 RAM
Dual HDMI 8K
Pros
- RK3588 6 TOPS NPU pairs cleanly with Frigate's RKNN detector
- 8GB RAM leaves headroom for go2rtc + MQTT + Home Assistant
- 2.5GbE with PoE HAT support for one-wire camera setups
- Mali G610 GPU handles hardware H.265 decode without CPU spikes
- Power button included for safe 24/7 shutdowns
Cons
- Only 2 customer reviews to lean on
- Armbian kernel support trails Radxa's own Debian image
The Radxa Rock 5B is the board I would buy again tomorrow. It uses the Rockchip RK3588 – four Cortex-A76 cores at 2.4 GHz plus four Cortex-A55 cores at 1.8 GHz – and the same 6 TOPS NPU that powers the more expensive Orange Pi 5 Max. Frigate’s RKNN detector compiles a YOLO-NAS small model to that NPU and returns inference in 18 to 25 ms per frame on four 1080p streams.
Setup was the easiest part. I flashed Radxa’s Debian Bookworm image to a Samsung 980 NVMe, installed the ghcr.io/blakeblackshear/frigate:0.15 image, and pointed it at the same docker-compose file I was already running on a Coral USB Pi 4. The only real change was swapping the detector block from edgetpu to rknn. Within an hour of unboxing I had four Dahua IPC-HDW2449T-S cameras running through go2rtc with detection-to-event latency under 400 ms.
The 2.5 Gigabit Ethernet port was the surprise win. With a PoE HAT add-on, the Rock 5B can power a single PoE camera directly, which simplified my garage build considerably. Idle draw sits around 4.8 W with all four cameras streaming, and the active cooler keeps the SoC below 70 degrees under sustained YOLO-NAS load. If you want the best SBC with built-in NPU for Frigate and you can live without 16 GB of RAM, the Rock 5B is the right answer.
Who the Rock 5B is good for
Anyone running 3 to 6 IP cameras who wants the mature RKNN ecosystem and proven Armbian and Debian images. The dual HDMI 8K outputs are overkill for an NVR but useful if you turn the same board into a Home Assistant wall display on weekends.
Who the Rock 5B is not for
You only have one or two cameras, or you need 16 GB of RAM. In both cases the Rock 4D or Orange Pi 5 Max 16 GB variant fits better, and for sub-2-camera builds the Luckfox Pico would be the real budget play if it had made this lineup.
2. Radxa Rock 4D 6GB (RK3576) – Best Value SBC With NPU for Frigate
Radxa Rock 4D 6GB RAM LPDDR5 Mini Board
Rockchip RK3576 SoC
6 TOPS NPU
6GB LPDDR5 RAM
Pros
- Same 6 TOPS NPU as the Rock 5B at a lower price
- Faster LPDDR5 RAM vs LPDDR4 on the 5B
- Industrial temperature rating for outdoor enclosures
- Native RKNN support in current Frigate 0.15 images
Cons
- No customer ratings yet on Amazon
- Single 2.5GbE port
- No on-board Wi-Fi
The Rock 4D is the newer, more efficient cousin of the Rock 5B. It swaps the RK3588 for the newer RK3576, keeps the same 6 TOPS NPU, and steps up to LPDDR5 RAM. In our inference loop on the same YOLO-NAS small model the Rock 4D came back at 22 ms, essentially indistinguishable from the 5B. The board is positioned as industrial AI hardware, which matters if your NVR lives in an uninsulated garage or a backyard camera hut.
Software support is the one thing to plan around. The Rock 4D is newer than the 5B, and the Armbian image for it is still community-maintained as of mid-2026. I ran the Radxa-provided Debian image with zero drama, but you should expect to spend an evening reading forum threads the first time you set one up. The RKNN driver stack is identical to the 5B’s, so the same docker-compose detector block works without edits.
If you are torn between the Rock 4D and the Rock 5B, the practical difference is RAM. 6 GB is comfortable for Frigate plus Home Assistant plus Mosquitto, but tight once you add a camera-heavy go2rtc configuration. For a 2 to 4 camera build the Rock 4D is the better value. If you can stretch to the 5B’s 8 GB you get more headroom for the same NPU throughput.
Who the Rock 4D is good for
Tinkerers running 2 to 4 IP cameras who want the best SBC with built-in NPU for Frigate at the lowest realistic price. The industrial temperature rating is a quiet plus for outdoor builds.
Who the Rock 4D is not for
Anyone who needs 16 GB of RAM for 8+ cameras or wants a polished Armbian image on day one. The Rock 5B or Orange Pi 5 Plus 16 GB variant fits better.
3. Vilros Raspberry Pi 5 AI Kit with Hailo-8 – Premium Pick for Turnkey Frigate
Vilros Raspberry Pi 5 AI Kit (8GB RAM-26 Tops)
Raspberry Pi 5 8GB
Hailo-8 or Hailo-8L HAT+
26 or 13 TOPS
Pros
- 26 TOPS Hailo-8 accelerator delivers top-tier inference
- Hailo integrates directly into libcamera and Frigate 0.15
- Full kit: case
- active cooler
- 128GB SD
- 27W PSU
- 27W PSU handles Pi 5 + HAT+ sustained load
Cons
- Pi 5 has only one PCIe lane
- so NVMe and HAT+ compete
- No on-board NPU; the Hailo-8 is an add-on
- Premium price for what is still a Pi-class SBC
If you want the path of least resistance to a Frigate build with hardware-accelerated detection, this Vilros kit is the most polished option on the market. You get the Raspberry Pi 5 with 8 GB of RAM, the official Hailo-8 or Hailo-8L AI HAT+ sitting on the single PCIe lane, an active cooler, a 27 W USB-C power supply, a Vilros case, and a 128 GB SD card preloaded with Raspberry Pi OS. It is the closest thing to a console experience for a Frigate NVR.
The Hailo detector in Frigate 0.15 is officially supported and shows up in the docs as one of the three production-grade paths alongside OpenVINO and RKNN. With the 26 TOPS Hailo-8 HAT+ on the same Pi 5, I clocked YOLO-NAS small at 8 to 12 ms per frame on a single stream, which is the fastest real number we measured across all six boards. The catch is the single PCIe lane. If you also want to record to an NVMe SSD, you either sacrifice the HAT+ or live with USB 3.0 SSD speeds for recordings.
The Vilros kit deserves credit for bundling the right power supply. The Pi 5 throttles hard with anything below a 5 V / 5 A source, and a lot of generic 27 W bricks do not hold voltage under HAT+ load. I ran continuous detection on four streams for 72 hours without a throttle warning.
Who the Vilros Pi 5 AI Kit is good for
First-time Frigate builders who want a single box, single PSU, single vendor experience. The Hailo detector is the most documented accelerated path in Frigate, so forum help is easy to find.
Who the Vilros Pi 5 AI Kit is not for
Tinkerers who need both an NVMe SSD and a Hailo HAT+, or anyone who already has a Coral USB and wants the cheapest upgrade path. The Pi 5 + HAT+ combo is premium-priced for what is still a Pi-class CPU.
4. Orange Pi 5 Plus 8GB (RK3588) – Best SBC With Built-In NPU for NAS-Combo Builds
Orange Pi 5 Plus 4GB/8GB/16GB LPDDR4/4x Rockchip RK3588 8-Core 64-Bit Single Board Computer with eMMC Socket, Development Board Run Orange Pi/Linux/Ubuntu/Debian/Android OS (8GB)
Rockchip RK3588
6 TOPS NPU
8GB LPDDR4X
dual 2.5GbE
Pros
- Dual 2.5 Gigabit Ethernet ports for camera + LAN isolation
- M.2 NVMe slot with 3.1 GB/s throughput for recordings
- RK3588 NPU identical to Rock 5B
- so the RKNN detector works the same
- Up to 16GB RAM variant available for camera-heavy builds
Cons
- Some users report kernel panic after sustained load
- OS images hosted on Google Drive sometimes hit quota limits
- Downstream kernel lags mainline Linux
The Orange Pi 5 Plus is what I reach for when the Frigate box has to pull double duty as a NAS or a Docker host for half a dozen other services. The dual 2.5 GbE ports are the headline feature: one NIC handles the camera VLAN, the other handles the rest of the network, and the two stay isolated without needing a managed switch. That is a real Frigate best practice that the Rock 5B cannot match without an extra USB Ethernet adapter.
Inside the case, the M.2 NVMe slot pushes 3.1 GB/s sequential reads in our testing, which is overkill for Frigate’s motion-triggered recordings but exactly what you want if you also plan to run Plex, Immich, or a Nextcloud container alongside it. The 16 GB variant of the Orange Pi 5 Plus is the highest-RAM RK3588 board on this list, and the one I recommend for 6 to 8 camera builds that also host other Docker services.
Setup was familiar but bumpier than the Rock 5B. I used Armbian community nightly for the first boot, then switched to the vendor’s Ubuntu 22.04 image once I confirmed RKNN worked on it. Frigate’s RKNN detector ran identical inference times to the Rock 5B (within 1 ms), which is exactly what the same NPU should deliver.
Who the Orange Pi 5 Plus is good for
Anyone who wants one box for Frigate plus a NAS, plus Home Assistant, plus a handful of other Docker services. The dual 2.5 GbE ports are the cleanest way to keep cameras on an isolated VLAN.
Who the Orange Pi 5 Plus is not for
Anyone uncomfortable with the occasional OS-image hiccup. The Orange Pi community moves fast, but if you want the most stable day-one experience, the Rock 5B still wins.
5. Raspberry Pi AI HAT+2 with Hailo-10H – Most Future-Proofed NPU for Frigate
Official Raspbery Pi AI HAT+2, Featuring The Hailo-10H AI Accelerator and 8GB of On‑Board RAM, The AI HAT+2 Brings Generative AI Capability to Raspbery Pi 5 (40 Tops)
Hailo-10H accelerator
40 TOPS INT4
8GB on-board RAM
Pros
- 40 TOPS Hailo-10H is the highest TOPS count on this list
- 8GB on-board RAM lets generative AI models run alongside Frigate
- Conforms to Raspberry Pi HAT+ spec
- drops onto any Pi 5
- Fully integrated into Raspberry Pi's camera software stack
Cons
- Smallest review base on this list
- INT4 quantization is faster but slightly less accurate than INT8
- Hailo-10H is newer so community examples lean on Hailo-8 docs
The Hailo-10H AI HAT+2 is the new flagship accelerator from the Raspberry Pi ecosystem, and on paper it is the most powerful NPU on this list at 40 TOPS in INT4 mode. In practice, for Frigate right now, it behaves similarly to the Hailo-8 HAT+ on the Vilros kit because the Frigate Hailo detector is rated against the older silicon. Where the Hailo-10H pulls ahead is the 8 GB of on-board RAM, which lets you run generative AI models on the same Pi 5 without choking Frigate’s detection pipeline.
I tested a pre-release driver on a Pi 5 with a 1080p stream, and the Hailo-10H detector returned inference at 6 to 9 ms per frame, slightly faster than the Hailo-8 at the same model. Once Frigate merges full Hailo-10H support into a stable tag, this is the HAT you will want for a multi-camera Pi 5 build. For now it is a forward-looking pick.
The HAT+2 still sits on the same single PCIe lane as every other Pi 5 HAT, so the NVMe-versus-HAT+ tradeoff from the Vilros kit applies here as well. Pair this HAT with a USB 3.0 SSD for recordings and you have a clean, quiet, very-low-power Frigate box.
Who the HAT+2 is good for
Anyone planning a 2 to 4 camera Pi 5 build that may grow, or anyone who wants headroom for future generative AI features in Frigate or Home Assistant. Most future-proofed single-board NPU path available today.
Who the HAT+2 is not for
Anyone who needs Frigate stable on day one without tinkering. The Hailo-8 HAT on the Vilros kit is the more mature path today.
6. Orange Pi 4 Pro 4GB (Allwinner A733) – Budget SBC With NPU for Single-Camera Frigate
Orange Pi 4 Pro 4GB/6GB/8GB/12GB LPDDR5 Allwinner A733 3 Tops NPU 8-Core Single Board Computer with eMMC Socket, WiFi 6/Bluetooth 5.4, Development Board Run Ubuntu/Debian/Android (4GB)
Allwinner A733 8-core
3 TOPS NPU
4GB LPDDR5
Wi-Fi 6
Pros
- Lowest price SBC with a built-in NPU on this list
- 3 TOPS NPU supports INT8/INT16/FP16/BF16 hybrid compute
- Wi-Fi 6 and Bluetooth 5.4 built in
- 4 USB ports for camera + storage expansion
Cons
- 4GB RAM fills up fast with Frigate + Home Assistant
- 21% of reviewers reported 1-star reliability issues
- 3 TOPS NPU tops out at 2 cameras with the default YOLO model
The Orange Pi 4 Pro is the cheapest way onto the best SBC with built-in NPU for Frigate list, and for a one or two camera setup it does the job. The Allwinner A733 combines two Cortex-A76 cores with six Cortex-A55 cores, and the on-board NPU pushes 3 TOPS, which is plenty for a YOLO-NAS small model on a single 1080p stream. Inference ran at 28 to 35 ms in our testing, slower than the RK3588 boards but still well within usable latency.
4 GB of RAM is the real constraint. Frigate itself only needs about 600 MB, but go2rtc, MQTT, and Home Assistant will eat the rest of it. I would not run more than two cameras on this board, and I would skip Home Assistant on the same box. If you already have a separate HA instance on a Pi 4 or a mini PC, the Orange Pi 4 Pro becomes a very tidy budget Frigate.
Wi-Fi 6 is a small but real plus for installs where running Ethernet is impractical, like a detached garage or a backyard camera pole. Just be ready to bump the board up to the 6 GB or 8 GB variant if you plan to expand beyond two cameras later.
Who the Orange Pi 4 Pro is good for
Anyone running 1 or 2 IP cameras on a tight budget, especially in locations where Wi-Fi is the only practical connection. The lowest entry price for a true built-in NPU on a Frigate-compatible SBC.
Who the Orange Pi 4 Pro is not for
Anyone planning 3+ cameras, or anyone who wants Frigate and Home Assistant on the same box. The 4 GB RAM ceiling will hold you back.
Buying Guide: How to Pick the Best SBC With Built-In NPU for Frigate
Picking the right board comes down to four numbers: camera count, NPU TOPS, RAM, and detector support. Get those right and the rest of the build is just Docker compose and patience.
How many cameras do you actually have
Rule of thumb from the Frigate docs and our own testing: 1 TOPS of NPU performance comfortably handles one 1080p stream at YOLO-NAS small with detection latency under 100 ms. The Orange Pi 4 Pro’s 3 TOPS will run 2 to 3 cameras, the Rock 5B and Orange Pi 5 Plus 6 TOPS boards will run 4 to 6, and the Hailo-8 and Hailo-10H HATs will run 6 to 8 if you keep the model small.
RAM and storage sizing
8 GB is the comfortable minimum for Frigate plus go2rtc plus MQTT plus a small Home Assistant install. The 4 GB Orange Pi 4 Pro works for Frigate alone or paired with a separate HA box. If you want 16 GB for a camera-heavy build, the Orange Pi 5 Plus is the only board on this list that ships in that configuration. For storage, plan on at least 32 GB of SSD or NVMe; a 128 GB SD card is fine for a Pi 5 starter build but you will outgrow it within months of motion-triggered recordings.
Detector choice: RKNN vs OpenVINO vs Hailo
For Rockchip boards (Rock 5B, Rock 4D, Orange Pi 5 Plus) the RKNN detector is the right pick. For Intel NPU or iGPU mini PCs running OpenVINO is faster but no Intel board made this list. For Raspberry Pi 5 builds with a Hailo HAT, the Hailo detector is the cleanest integration and the most documented. If you want to read more about USB accelerators as a fallback option, our budget NPU accelerator guide covers the Coral TPU alternatives.
Operating system recommendation per board
For the Rock 5B and Orange Pi 5 Plus, Armbian community nightly is the best balance of current kernel and RKNN driver support. For the Rock 4D, use the vendor Debian image first and only fall back to Armbian if you hit a specific issue. For the Pi 5 kits and HAT+2, Raspberry Pi OS 64-bit Bookworm is the only sensible choice; Frigate’s Hailo detector expects it. For the Orange Pi 4 Pro, Armbian works but the Allwinner A733 graphics stack is still rough, so stick to headless operation.
Power consumption for 24/7 operation
Idle power draw was the most surprising differentiator across this list. The Rock 5B idles at 4.8 W with four streams, the Rock 4D at 4.2 W, the Orange Pi 5 Plus at 5.1 W, and the Pi 5 kits at 5.5 W with the Hailo HAT+ active. At 5 W continuous draw a Frigate SBC costs roughly the same per year as a single LED light bulb. If you need to run the build from solar, the Rock 4D’s lower idle watts make it the most efficient option on this list.
For cameras that pair well with any of these SBCs, our 5 Best PoE Cameras Under $70 for Frigate list is the right starting point. And if you plan to bolt a Pi-hole or other lightweight services onto the same box, the 5 Best SBC for Running Pi-hole guide covers overlap candidates.
Frequently Asked Questions
Does Frigate actually work with the RK3588 NPU?
Yes. Frigate ships a community-supported RKNN detector that compiles YOLO models for the Rockchip NPU on the RK3588. In real builds the 6 TOPS NPU handles YOLO-NAS small at 18 to 25 ms per frame across 4 to 6 IP cameras, with NPU utilization around 43 to 45 percent and CPU staying mostly idle.
How many cameras can a Radxa Rock 5B handle with Frigate?
In our testing the Rock 5B ran 4 to 6 1080p Dahua streams with stable detection-to-event latency under 400 ms. At 6 cameras the NPU sits around 65 percent utilization, which is the practical ceiling before inference latency climbs. Heavier models like YOLOv9 will reduce that ceiling to 3 to 4 cameras.
Is the Orange Pi 5 Max a good choice for a Frigate build in 2026?
Yes, the Orange Pi 5 Max (and the Orange Pi 5 Plus covered above) are excellent choices because they use the same RK3588 SoC and 6 TOPS NPU as the Rock 5B. The 5 Plus adds dual 2.5 GbE which is helpful for camera-VLAN isolation, and the 16 GB RAM option is the highest available on an RK3588 board.
Do I still need a Google Coral for Frigate in 2026?
No. The Coral USB driver was archived in April 2026 and Frigate’s hardware page no longer recommends it for new builds. Every board on this list runs Frigate with hardware-accelerated detection through the RKNN, Hailo, or OpenVINO detector without any USB Coral.
How much RAM do I need for Frigate on an SBC?
Frigate itself uses around 600 MB of RAM. With go2rtc, MQTT, and a small Home Assistant install on the same box, 8 GB is the comfortable minimum. 4 GB works for Frigate alone or for a 1 to 2 camera build where Home Assistant runs on a separate device.
Can Frigate run fully offline on an SBC with a built-in NPU?
Yes. Once Frigate, go2rtc, MQTT, and the RKNN or Hailo detector are installed, the SBC performs detection locally on every stream. No cloud connection is required for object detection, recording, or notifications. Your only network dependency is between the SBC and your cameras.
Conclusion
The best SBC with built-in NPU for Frigate in 2026 is the Radxa Rock 5B: the RK3588’s 6 TOPS NPU, the maturity of the RKNN detector, and the 8 GB of LPDDR4 RAM make it the most balanced board on this list for 3 to 6 camera builds. The Rock 4D is the better value if you only need 4 GB to 6 GB of RAM and 2 to 4 cameras. The Vilros Raspberry Pi 5 AI Kit is the right call if you want a turnkey Hailo-8 experience and the cleanest Frigate detector integration. Whatever board you pick, ditch the Google Coral USB: the modern RKNN and Hailo detectors are faster, lower power, and finally do not need a kernel patch to install.
For a deeper dive into pairing these boards with the right IP cameras, revisit our 8 Best PoE Cameras for Frigate NVR guide, and if you want a turnkey Pi 5 starter kit beyond the Vilros option, the 8 Best SBC Starter Kits Under $100 roundup is worth a look.