Best Coral Alternatives for Frigate NVR 2026: 6 Picks

Disclosure: As an Amazon Associate, The Gear Nest earns from qualifying purchases. This adds no cost to you.

Frigate’s 2026 documentation flipped a switch that shook up every DIY NVR build: the Google Coral is no longer recommended for new installations. If you’re planning a Home Assistant-connected NVR this year, the conversation around Coral alternatives for Frigate now revolves around Hailo, Intel, and — for anyone with existing hardware — a handful of legacy Coral options that still work. Here are six real, buyable Coral alternatives for Frigate NVR builds in 2026, matched to actual host hardware.

Prices accurate at time of writing and subject to change.

Comparison Table: 6 Coral Alternatives for Frigate at a Glance

Product	Best For	Key Spec	Price Tier	Rating	Buy Link
Raspberry Pi AI HAT+ 13 TOPS	Pi 5 Frigate builds	Hailo-8L, 13 TOPS	$	4.6/5	Check Price on Amazon
Raspberry Pi AI HAT+ 26 TOPS	Pi 5, 4-10 cameras	Hailo-8, 26 TOPS	$$	4.7/5	Check Price on Amazon
Hailo-8 M.2 Module (bare)	x86 mini PCs	26 TOPS, 2.5W	$$	4.5/5	Check Price on Amazon
Intel Arc A380 (6GB)	Detect + transcode combo	6GB GDDR6, PCIe 4.0	$$	4.4/5	Check Price on Amazon
Google Coral USB Accelerator	Existing Pi 4 / low-power builds	Edge TPU, 4 TOPS	$	4.3/5	Check Price on Amazon
Coral M.2 Dual Edge TPU	Existing M.2 E-key slots	2x Edge TPU, 8 TOPS	$	4.2/5	Check Price on Amazon

Raspberry Pi AI HAT+ (13 TOPS, Hailo-8L)

The cleanest replacement for a Coral on a Raspberry Pi 5, using the exact silicon Frigate now recommends.

Why it makes the list: The AI HAT+ 13 TOPS is the current, in-production version of what used to ship as the Raspberry Pi AI Kit. It uses the same Hailo-8L chip, but soldered directly onto the HAT+ board instead of a separate M.2 module — which means one less connector to seat and one less thing that can rattle loose. The AI Kit is no longer in production, and Raspberry Pi’s own documentation now points buyers at the AI HAT+ line. At $70 it’s the lowest-friction “buy this instead of a new Coral” recommendation for a Pi 5 Frigate build.

Key specs:

• 13 TOPS INT8 inference on the Hailo-8L accelerator
• Connects via the Raspberry Pi 5’s PCIe Gen 3 interface
• Fully integrated into the rpicam-apps camera software stack — auto-detected on up-to-date Raspberry Pi OS
• Conforms to the Raspberry Pi HAT+ specification, with a 16 mm stacking header, spacers, and screws included for fitting alongside the Pi 5 Active Cooler
• Operating temperature 0-50°C ambient
• Raspberry Pi lists it as in production until at least January 2030
• $70 list price

Watch out for: This board only works with a Raspberry Pi 5 — it will not run on a Pi 4 or earlier. If you’ve already bought models specifically compiled for the Hailo-8 (26 TOPS), some of those won’t run on the 13 TOPS Hailo-8L, though lower-performance alternatives are usually available in the Hailo Model Zoo. And if you’ve been eyeing the older AI Kit on a marketplace listing, know that Raspberry Pi has moved on and long-term firmware/model support will follow the HAT+ line.

Raspberry Pi AI HAT+ (26 TOPS, Hailo-8)

Twice the neural compute of the 13 TOPS variant for larger camera counts and multiple parallel models.

Why it makes the list: The 26 TOPS AI HAT+ swaps in the full Hailo-8 chip — twice the on-die compute resources of the Hailo-8L — for an extra $40. It automatically switches the Pi 5’s PCIe interface to Gen 3.0 mode to feed the accelerator’s full bandwidth, and it handles multiple concurrent neural network models without dropping frames. For Frigate builds targeting 4-10 cameras, or setups where you want object detection running in parallel with a second model (like pose estimation or face recognition), this is the pick.

Key specs:

• 26 TOPS INT8 inference on the Hailo-8 chip
• Accelerator soldered directly onto the HAT+ PCB (no separate M.2 module)
• Auto-switches the Pi 5’s PCIe port to Gen 3.0 mode for full bandwidth
• Handles multiple simultaneous neural network models
• Independent testing puts Hailo-8 at roughly 10.4 TOPS per watt versus Coral’s 2 TOPS per watt, at about 2.5W sustained draw
• Ships with the 16 mm stacking header, spacers, and screws needed to fit alongside the Pi 5 Active Cooler
• $110 list price

Watch out for: Like the 13 TOPS variant, this needs a Raspberry Pi 5. It also occupies the Pi 5’s PCIe port, so if you were planning to run a Pi 5 NVR with an NVMe SSD in the same slot you’ll need a PCIe adapter or a workaround. The AI HAT+ (13 or 26 TOPS) does not ship with a heatsink — for a 24/7 NVR that pegs the accelerator hard, plan on the Pi 5 Active Cooler plus adequate case ventilation.

Hailo-8 M.2 AI Acceleration Module (bare)

The x86 mini-PC equivalent of the AI HAT+ — same 26 TOPS silicon, no HAT.

Why it makes the list: If your Frigate host is a Beelink, MinisForum, N100 mini PC, or any small x86 box with a spare M.2 slot, the bare Hailo-8 M.2 module is the direct drop-in for what a Coral M.2 used to do. It ships in M-key, B+M-key, and A+E-key variants so it can slot into a WiFi socket or a proper storage slot depending on the host. Unlike Coral, all the memory sits on-chip, so there’s no external DRAM to fail and no supply-chain dependency on separate memory parts.

Key specs:

• 26 TOPS INT8 inference on the Hailo-8 processor
• Typical power consumption ~2.5W
• PCIe Gen 3 x2 or x4 (host dependent)
• Available in M, B+M, and A+E key M.2 variants (2242 base, breakable to 2280)
• Linux and Windows support via HailoRT
• Supports TensorFlow, TensorFlow Lite, Keras, PyTorch, and ONNX model formats

Watch out for: Unlike the Pi AI HAT+, this is a bare module — no HAT, no cables, no heatsink in the box. You’ll need to install the HailoRT driver on the host before the Frigate container can see /dev/hailo0; on Debian and Ubuntu x86 that means running Hailo’s installation script, but DKMS handles subsequent kernel upgrades cleanly. Most importantly, verify your host’s M.2 slot key type before ordering: M, B+M, and A+E variants are not interchangeable.

Intel Arc A380 (6GB, ASRock Challenger ITX)

The one card that can do object detection AND camera transcoding on a single PCIe slot.

Why it makes the list: Every accelerator on this list handles inference. Only one — a discrete GPU — also handles video decode for the H.264/H.265 camera streams that Frigate has to process before it can even detect anything. Community benchmarks on Frigate 0.15 show the ASRock Arc A380 Challenger 6GB running YOLO-NAS-s-320.onnx via OpenVINO on the GPU, with the card passed through to a Proxmox VM. The A380 also has unrestricted concurrent hardware encode sessions — useful when Frigate needs to re-encode streams for multiple substreams simultaneously — and industry-first AV1 hardware encode and decode.

Key specs:

• 6GB GDDR6 memory on a 96-bit bus
• PCI Express 4.0 host interface
• Hardware AV1 encode and decode (plus H.264 and H.265)
• Unrestricted concurrent hardware encoding sessions
• Confirmed to run YOLO-NAS via Frigate’s OpenVINO GPU detector
• Low-profile ITX variants (ASRock Challenger) fit small form factor Proxmox boxes

Watch out for: The A380 needs a motherboard with Resizable BAR enabled to perform well — older boards without ReBAR support have caused problems for Frigate community members. Idle power draw is meaningfully higher than a Hailo module or an Intel iGPU, so if you don’t also want the transcode capability, this is overkill. And if you’re running it under Proxmox with PCIe passthrough, expect to update packages inside the Frigate container to smooth out OpenVINO compatibility.

Google Coral USB Accelerator

Still supported, still shipping, still the simplest install — but no longer the default recommendation.

Why it makes the list: Frigate hasn’t dropped Coral support and won’t anytime soon. If you already own a Coral USB stick, or your host is a Raspberry Pi 4 where USB 3 is the only viable interface, or you genuinely need the lowest-power option, the Coral USB Accelerator remains a legitimate pick. Included on this list so readers with existing Coral hardware know when it’s still okay to keep using it — and when it isn’t.

Key specs:

• 4 TOPS INT8 inference on Google’s Edge TPU ASIC
• 2W typical power draw (roughly 2 TOPS per watt)
• USB 3.0 interface — no host driver required beyond the Edge TPU runtime
• Runs MobileNet v2 at approximately 400 FPS
• Compatible with x86-64, Armv7, and Armv8 Linux hosts on Debian 10+ or Ubuntu 18.04+
• Frigate inference times land in the 5-10 ms per frame range

Watch out for: Frigate’s 2026 documentation explicitly says Coral is no longer recommended for new installs, so buying one today only makes sense if you have a specific existing-hardware constraint. Google’s official datasheet warns that the device gets very hot to the touch at max clock frequency, with a burn injury risk — keep it out of reach or run the reduced-frequency mode. Only one model can run on a single Coral at a time, so multi-model pipelines are limited. And stock is intermittent: Mouser and DigiKey have listed it as backordered on and off through 2026.

Google Coral M.2 Dual Edge TPU (A+E Key)

A legacy pick that still earns a spot on existing Coral-driver builds with the right M.2 E-key slot.

Why it makes the list: This module packs two Edge TPUs onto a single M.2 2230 E-key card for 8 TOPS total. For readers who already have a working Coral toolchain, a compatible M.2 E-key slot with both PCIe x1 lanes wired, and don’t want to switch drivers, the Dual Edge TPU can double a single-Coral install’s throughput. Frigate 0.14 and later support multiple Coral detectors in the same config, so both TPUs can be worked in parallel.

Key specs:

• 8 TOPS total (2 x 4 TOPS Google Edge TPU coprocessors)
• ~2W per TPU (roughly 2 TOPS per watt)
• 22 x 30 x 2.8 mm M.2 2230 E-key form factor
• Two PCIe Gen 2 x1 interfaces (one per TPU)
• Debian Linux and Windows 10 support
• Per-TPU configurable thermal shutdown with built-in temperature sensor

Watch out for: The single biggest gotcha with this card is the M.2 spec itself. E-key sockets are supposed to provide two PCIe x1 lanes, but most consumer motherboards only wire one. If that’s your case, only one of the two Edge TPUs will be usable and you’ve effectively paid double for a single Coral. Check your host’s M.2 E-key slot documentation before ordering. Peak current draw can hit 3A per TPU under load, so the slot also needs to supply the power. Like the Coral USB stick, this is now a legacy pick — reserve it for existing Coral-equipped hosts rather than fresh builds.

How to Choose a Coral Alternative for Frigate in 2026

Before picking hardware, count your cameras and your target detect FPS (5 fps per camera is a common baseline). Frigate’s own capacity model is cameras ≈ (1000 ÷ inference_ms) ÷ detect_fps, then de-rate roughly 25% for real-world overhead. A 10 ms detector at 5 fps gives you around 20 camera-streams of headroom on paper — plenty for most home setups.

From there, three questions decide the pick:

1. What’s the host? If it’s a Raspberry Pi 5, the AI HAT+ (13 or 26 TOPS) is the cleanest answer. If it’s an x86 mini PC with a free M.2 slot, the bare Hailo-8 M.2 module is the equivalent. If it’s a bigger box with a PCIe x16 slot and you also want camera transcoding on the same silicon, an Intel Arc A380 pulls double duty.
2. How many cameras? One to three cameras run happily on the 13 TOPS Hailo-8L, a Coral USB, or a Coral Dual TPU. Four to ten cameras want the 26 TOPS Hailo-8, or an Intel Arc A380 if you also need decode offload.
3. What’s already in the drawer? If you already own a Coral (USB or M.2), it still works, and Frigate will keep supporting it. Don’t rip it out on principle — just don’t plan a new build around it.

Not sure which cameras to feed into the setup? Our roundup of the best PoE cameras for Frigate NVR covers models that work well as the input side of this whole equation, and if you’re still speccing the box itself, our guide to the best mini PCs for Proxmox lists hosts with the M.2 or PCIe slots these accelerators need.

Frequently Asked Questions

Is the Google Coral discontinued?

Not officially, but the Coral website hasn’t been meaningfully updated in years and Frigate’s 2026 hardware documentation states the Coral is no longer recommended for new Frigate installations, except in very low-power builds or hardware that can’t fit an alternative. Frigate continues to support Coral for as long as practicably possible given how power-efficient it is, but the project now steers new users toward Hailo, Intel iGPU/NPU via OpenVINO, or a supported GPU.

Do I need an AI accelerator for Frigate?

Not always. Frigate now works well with a modern Intel iGPU via OpenVINO, and a supported NVIDIA GPU can also handle detection. The practical rule is to buy a dedicated detector like the Hailo AI HAT+ only after your host’s iGPU or NPU can’t keep up with your camera count. That said, an accelerator massively reduces CPU load and is still the right answer for a Raspberry Pi 5 build where the CPU is doing everything else too.

What is the best Coral alternative for a Raspberry Pi 5?

The Raspberry Pi AI HAT+ 13 TOPS ($70) for one to three cameras, or the AI HAT+ 26 TOPS ($110) for four to ten cameras or multi-model workloads. Both are officially supported first-class Frigate detectors, both use rpicam-apps auto-detection, and both are backed by Raspberry Pi’s own supply commitment through at least January 2030.

Will my old Coral USB stick still work with Frigate 0.16?

Yes. Coral EdgeTPU is still fully supported and compatible with Frigate 0.16 in both USB and M.2/PCIe formats. Inference times remain in the 5-10 ms per frame range. Multiple Coral devices can be configured as separate detectors in the same Frigate install if you want to scale a Coral-based deployment without switching to a new accelerator.

Do I need a separate GPU for video decoding?

Detection and decoding are separate jobs. The accelerator runs object inference; turning H.264 or H.265 camera streams into frames is a separate task for the host CPU, iGPU, or a discrete GPU. A fast detector on a host that can’t decode your streams will still drop frames. This is exactly why the Intel Arc A380 stands out on this list — it’s the one option that handles both inference and hardware transcoding on the same card.

Conclusion

The 2026 answer to “what Coral should I buy for Frigate?” is usually “buy a Hailo instead.” Across the six Coral alternatives for Frigate on this list, the right pick tracks the host hardware more than anything else. For a Raspberry Pi 5, that’s the AI HAT+ 13 TOPS at $70 for small setups, or the 26 TOPS variant at $110 for anything more demanding. For an x86 mini PC with a spare M.2 slot, the bare Hailo-8 module gives you the same silicon in a form factor that fits Beelink and MinisForum-class boxes. If you’re building on a machine with a real PCIe x16 slot and want your detector to also transcode camera streams, the Intel Arc A380 handles both. And if you already own a Coral, keep using it — the two Coral picks on this list exist precisely for people whose hardware, budget, or driver stack makes swapping impractical right now.