Buyer’s guide

Best Edge AI Chips for Drones

For most drone programs, the best edge AI chips in 2026 are NVIDIA’s Jetson Orin family for mid-size aircraft, Qualcomm’s Dragonwing QRB5165 and QCS8550 for small, power-limited drones, and Hailo accelerators when you need to add AI to an existing processor. NVIDIA Jetson Thor sits at the top end for large platforms that can carry 40 W or more of compute.

TOPS figures are the vendors’ own and use different precisions, so treat them as a rough guide and benchmark your own models.

Method

How we chose

01Current, orderable partsEvery chip or module is listed on the vendor’s site today.

02Drone or robotics fitThe vendor names drones or robotics as a target, or the part ships in drone hardware.

03Published performanceEach entry has a vendor-stated AI figure.

04Range of sizesWe cover everything from 2.5 W accelerators to 130 W modules.

At a glance

Quick comparison

Chip or module	HQ	Best for	Notable (vendor figures)
NVIDIA Jetson AGX Thor	Santa Clara, CA	Large UAVs and heavy multimodal AI	T5000: 2070 FP4 TFLOPS (sparse), 40–130 W
NVIDIA Jetson AGX Orin	Santa Clara, CA	Mid-to-large drones with many sensors	Up to 275 TOPS, 15–60 W
NVIDIA Jetson Orin NX and Orin Nano	Santa Clara, CA	Small and mid-size drones on Jetson	34–157 TOPS, 7–25 W
Qualcomm Dragonwing QRB5165	San Diego, CA	Very small, power-limited drones	15 TOPS, Flight RB5 5G platform
Qualcomm Dragonwing QCS8550	San Diego, CA	Next-generation small-drone compute	48 INT8 TOPS
Qualcomm Dragonwing IQ-9075	San Diego, CA	Industrial-grade drones and robots	100 dense TOPS
Hailo-8 and Hailo-8L	Tel Aviv, Israel	Adding vision AI to a host CPU	26 and 13 TOPS; Hailo-8 at 2.5 W typical
Hailo-10H	Tel Aviv, Israel	Generative AI models at the edge	40 TOPS INT4, 20 TOPS INT8
Ambarella CV72S and CV7	Santa Clara, CA	Camera-centric drones	CV7: 2.5x CV5 AI performance
Rockchip RK3588	Fuzhou, China	Low-cost embedded Linux designs	6 TOPS NPU

NVIDIA Jetson AGX Thor

Best for

Large drones and robots running several AI models at once.

Jetson Thor is NVIDIA’s top embedded module, built on the Blackwell GPU. The Jetson T5000 lists 2070 FP4 TFLOPS (sparse), 128 GB of LPDDR5X and a 40–130 W power range. The smaller T4000 lists 1200 FP4 TFLOPS (sparse), 64 GB and 40–70 W. That power budget rules Thor out for small multirotors, but it suits large fixed-wing aircraft, VTOLs and ground vehicles that need vision-language models or dense sensor fusion. Rugged carrier boards are available from partners such as Connect Tech.

NVIDIA Jetson AGX Orin

Best for

Mid-to-large drones fusing many cameras and lidar.

Jetson AGX Orin 64GB is rated at 275 TOPS and AGX Orin 32GB at 248 TOPS, both in a 15–60 W range. An industrial version lists 241 TOPS at 15–75 W for harsher temperature and vibration conditions. AGX Orin has broad software support through JetPack and NVIDIA Isaac ROS, including GPU-accelerated visual SLAM and 3D mapping packages. That makes it a practical choice for inspection and mapping platforms that process several high-resolution camera or lidar streams on board and still need headroom for new models.

NVIDIA Jetson Orin NX and Orin Nano

Best for

Small and mid-size drones that want the Jetson software stack at lower power.

In NVIDIA’s Super configurations, Orin NX 16GB reaches 157 TOPS and Orin NX 8GB 117 TOPS, both at 10–25 W. Orin Nano 8GB reaches 67 TOPS and Orin Nano 4GB 34 TOPS at 7–15 W. These modules drop into drone-ready carriers such as the ARK Jetson PAB V3, which links the Jetson directly to a Pixhawk-standard flight controller. They share NVIDIA’s JetPack software with AGX Orin, so teams can prototype on a large module and fly a smaller one without changing tools.

Qualcomm Dragonwing QRB5165

Best for

Very small drones where every gram and watt counts.

The QRB5165 delivers 15 TOPS from Qualcomm’s fifth-generation AI Engine and adds an image signal processor with multi-camera support and a dedicated computer vision engine. It powers the Qualcomm Flight RB5 5G Platform, which supports up to seven cameras, Wi-Fi 6 and 5G for beyond-visual-line-of-sight work. ModalAI’s 16-gram VOXL 2, a combined flight controller and AI computer, also uses the QRB5165. It is a mature, well-documented choice for compact visual navigation and obstacle avoidance, with development kits available from Thundercomm and others.

Qualcomm Dragonwing QCS8550

Best for

Small drones that need several times the AI performance of the QRB5165.

The QCS8550 targets industrial drones, autonomous mobile robots and AI cameras, and adds Wi-Fi 7 connectivity. On Lantronix’s Open-Q 8550CS µSOM, launched in April 2026 for drone integration, it delivers 48 INT8 TOPS and 12 FP16 TOPS. Radxa and other module makers also offer QCS8550 boards, which gives buyers more than one supply option. For teams already on the QRB5165, it is a natural upgrade path that stays within the Qualcomm software stack and keeps the same multi-camera design approach.

Qualcomm Dragonwing IQ-9075

Best for

Industrial-grade drones and robots running larger models.

The IQ-9075 is an industrial-grade processor with an NPU rated at 100 dense TOPS, and Qualcomm says it can run models of about 13 billion parameters on device. Its evaluation kit pairs it with 36 GB of LPDDR5 and CAN-FD, PCIe and MIPI CSI interfaces. Qualcomm names UAVs, drones and agricultural robots as target uses. It sits between the QCS8550 and Jetson AGX Orin for teams that prefer the Qualcomm ecosystem but need far more AI headroom than a small-drone chip.

Hailo-8 and Hailo-8L

Best for

Adding vision AI to an existing processor without a new SoC.

Hailo-8 is an AI accelerator rated at 26 TOPS with 2.5 W typical power, and it keeps all its memory on the die, so no external DRAM is needed. Hailo-8L is a lower-performance version at 13 TOPS (INT8), used on the Raspberry Pi AI HAT+. Both come as chips or M.2 modules, so a drone with a capable host processor can add neural network inference over PCIe. Hailo’s compiler and runtime support TensorFlow, PyTorch and ONNX models, and its Model Zoo offers pre-optimized networks.

Hailo-10H

Best for

Running small language or vision-language models on the aircraft.

Hailo-10H is Hailo’s second-generation accelerator. It is rated at 40 TOPS INT4 and 20 TOPS INT8 at 2.5 W typical. Unlike Hailo-8, it has a direct DDR interface for LPDDR4/4X, so it can hold larger models such as LLMs and VLMs. It comes in chip and M.2 forms with industrial (-40°C to 85°C) and automotive (-40°C to 105°C) grades. The Raspberry Pi AI HAT+ 2 uses it for local language models, which makes it easy to evaluate before designing it into a drone.

Ambarella CV72S and CV7

Best for

Drones where the camera is the product.

Ambarella’s CV-series SoCs combine image processing, video encoding and CVflow AI acceleration on one chip, which saves weight and power in camera-heavy designs. The 5 nm CV72S lists drones among its target applications. The CV7, announced at CES 2026, is built on a 4 nm process and, per Ambarella, delivers 2.5 times the AI performance of the earlier CV5 with 20% less power while handling 8Kp60 video. Antigravity’s A1 drone uses the earlier CV5 for 8K imaging and on-device inference.

Rockchip RK3588

Best for

Cost-sensitive embedded Linux designs with light AI loads.

The RK3588 combines an eight-core CPU with a triple-core NPU rated at up to 6 TOPS. The NPU supports INT4, INT8, INT16 and FP16, and its cores can work alone or together, with models converted from TensorFlow, PyTorch and other frameworks. It is widely available on single-board computers and modules. Rockchip is a Chinese company based in Fuzhou, so programs with NDAA or FCC sourcing requirements should review component origin with their compliance team before they design it in.

FAQ

Common questions

How many TOPS does a drone need?

It depends on the models and cameras. Single-camera object detection can run on 13 to 26 TOPS accelerators like Hailo-8L or Hailo-8. Visual navigation with several cameras fits well on a 15 TOPS QRB5165. Multi-sensor fusion, lidar processing or language models push you toward Orin NX, AGX Orin or Thor. Always benchmark your own models.

Jetson or Qualcomm for a small drone?

Qualcomm Dragonwing parts tend to suit tight weight and power budgets, and they include strong camera pipelines. Jetson offers more GPU compute and the CUDA and Isaac ROS software ecosystem. ModalAI’s 16-gram VOXL 2 shows the Qualcomm approach; Jetson carriers such as ARK’s PAB V3 suit drones with more payload capacity.

Are TOPS numbers comparable across vendors?

Not directly. Vendors quote different precisions (INT4, INT8, FP16, FP4) and sometimes sparse figures. Hailo-10H, for example, is 40 TOPS at INT4 but 20 TOPS at INT8. Compare at the same precision and, better still, compare frames per second on your own network.

Does the chip affect NDAA compliance?

It can. NDAA sourcing rules (FY2020 §848 and FY2024 §1823-1826) focus on specific components and covered manufacturers, so the processor’s origin matters. FCC Covered List status is a separate question that applies to the finished product. Check both with your supplier and compliance team before you commit to a design.

See all 310 questions in the drone and robotics FAQ →Look up terms in the glossary →

Choosing and integrating

The board, thermal design and camera interfaces decide whether a chip works in the air. Aerora’s engineers design on Qualcomm, NVIDIA, Rockchip, Intel, TI, ST and Ambarella platforms; see its engineering capabilities or our list of the best edge AI integration partners for autonomous drones.

Sources (23)

Start a program

Ready to build what’s next?

Whether you’re developing autonomous aircraft, robotics, intelligent sensing systems or connected platforms, Aerora can help bring your vision to life.

Submit a design inquiry