Edge AI Chips
2,070 TFLOPS to 3.5W — the full edge AI chip landscape
Edge AI chips span five orders of magnitude in power consumption and compute. Choosing the right chip for a use case is not about picking the most powerful — it is about matching compute, power budget, memory bandwidth, and software ecosystem to the actual deployment context.
Chip comparison
Every major edge AI chip — verified specs
NVIDIA Jetson Thor (T5000)
Robotics / HumanoidsCompute
2,070 FP4 TFLOPS
Memory
128GB LPDDR5X
Power
40–130W
Price
$3,499 dev kit
Blackwell GPU, 14-core Arm Neoverse-V3AE. 7.5× compute vs Orin. Targets humanoid robots running GR00T N1.
Adopters: Boston Dynamics Atlas, Agility Digit, 1X NEO, Amazon Robotics, Meta
Best for: Humanoid robots, industrial manipulation, autonomous vehicles
NVIDIA Jetson AGX Orin
Current robotics standardCompute
275 TOPS INT8
Memory
64GB
Power
60W max
Price
$899 dev kit
Current deployed standard across AMRs, drones, and robotic arms. Well-supported ROS2 ecosystem.
Adopters: Widely deployed — Boston Dynamics Spot, AMR manufacturers, drone vendors
Best for: AMRs, drones, robotic arms, industrial inspection
Qualcomm Snapdragon X2 Elite
Laptops / WearablesCompute
80 TOPS NPU (Hexagon 6)
Memory
Up to 64GB LPDDR5X
Power
23W TDP (16% lower vs prior gen)
Price
OEM (not standalone)
16% lower power consumption versus prior generation. Primary competitor to Apple M-series for Windows AI laptops and ARM-based edge devices.
Adopters: Windows AI PCs, ARM edge computing devices
Best for: AI laptops, Windows on ARM, mid-power edge inference
Apple M4 Neural Engine
Consumer / On-device LLMCompute
38 TOPS Neural Engine, 16-core
Memory
16–32GB unified (configurable)
Power
~15–20W
Price
Apple Silicon devices
TSMC N3E. Unified memory architecture eliminates data copy overhead — critical for on-device LLM inference. Best performance-per-watt for Apple hardware. Runs Llama 3 8B at ~50 tok/s.
Adopters: MacBook, iPad Pro, iPhone 16 (A18)
Best for: On-device LLM inference, Apple ecosystem AI, personal device AI
Hailo-10H
Ultra-low power edgeCompute
40 TOPS
Memory
On-chip SRAM
Power
Under 3.5W
Price
Module pricing
Best power efficiency for edge inference. Runs 2B parameter models at 2.5W. Llama2-7B at 10 tok/s. Designed for always-on embedded use cases where power budget is the primary constraint.
Adopters: Embedded edge devices, IoT AI, edge servers
Best for: Always-on embedded AI, battery-powered edge devices, smart home
Intel Loihi 2
Research neuromorphicCompute
~1M neurons
Memory
128MB on-chip SRAM
Power
~100mW
Price
Research access only
Neuromorphic architecture — event-driven, sparse computation. Not production-ready. Relevant for understanding the long-term trajectory of ultra-low-power AI hardware.
Adopters: Intel Labs, academic research
Best for: Research — not for production deployment
Use-case guide
Which chip for which application
| Use case | Chip |
|---|---|
| Humanoid robot | Jetson Thor (T5000) |
| Industrial AMR / drone | Jetson AGX Orin |
| AI laptop / Windows ARM | Qualcomm Snapdragon X2 Elite |
| On-device LLM (Apple) | Apple M4 |
| Always-on wearable AI | Hailo-10H |
| Smart home edge AI | Hailo-10H |
Tradeoffs
Power vs performance — the real constraint
The TOPS or TFLOPS number is not what matters in isolation. The relevant figure is TOPS per watt — inference performance per unit of power consumed. A chip that runs at 3.5W continuously is deployable in a wearable or smart home device. A chip that requires 130W requires active cooling and a permanent power connection.
Wearables
Power budget: Under 5W
Hailo-10H
Continuous inference at 3.5W. Battery viability depends on total system power, not chip alone.
Edge servers / robots
Power budget: 20–130W
Jetson AGX Orin / Thor
Active cooling required for Thor at 130W. Most robotics platforms manage this with chassis design.
Laptops / handhelds
Power budget: 10–25W
Apple M4 / Qualcomm X2 Elite
The sweet spot for consumer on-device AI — enough compute for 7B parameter models, low enough for fanless designs.
Last updated: July 2026