Compare
Chips and edge compute
Compare chips and edge compute platforms for robotics: AI performance, power draw, price, availability.
These chips are just one layer among others: our physical AI stack pillar places them within the wider system.
Click any row to reveal details, sources and verification date.
| Name | Vendor | Country | Type | AI perf | Power (W) | Price (USD) | Availability |
|---|---|---|---|---|---|---|---|
| NVIDIA Jetson AGX Thor (T5000) | NVIDIA | US | GPU module | 2070 TFLOPS FP4 (sparse, Blackwell GPU) | 130 | $3,499 | GA since August 2025 (dev kit $3,499; production modules shipping) |
|
The de facto flagship compute for humanoid robots in 2026: 7.5x the AI compute of Orin, 128 GB of memory to run VLA models on-robot, adopted or evaluated by nearly every major humanoid maker including Boston Dynamics, Agility, Figure and OpenAI.
Memory : 128 GB LPDDR5XForm factor : System-on-module (Jetson T5000) + developer kit; 40-130 W envelopeApplications : humanoid, AMR, manipulation, healthcare, agriculture, visionUsed by : Boston Dynamics Atlas, Agility Robotics Digit, Figure, Amazon Robotics, Unitree H2 Plus, Caterpillar, 1X (evaluating)
Sources : NVIDIA Newsroom (2025-08-25)CNX Software (2025-08-19)NVIDIA (2026-07-09) Last verified 2026-07-09 |
|||||||
| NVIDIA Jetson AGX Orin | NVIDIA | US | GPU module | 275 TOPS INT8 (sparse; 138 TOPS dense) | 60 | $1,999 | GA since 2022; the installed-base reference for embedded robotics AI |
|
Four years after launch, Orin remains the workhorse of the robotics industry, powering everything from Unitree humanoids to Skydio drones; Thor supersedes it at the high end but Orin keeps the volume and the mature Isaac/ROS ecosystem.
Memory : 64 GB LPDDR5 (unified)Form factor : System-on-module + developer kit; 15-60 W configurableApplications : humanoid, AMR, drone, vision, industrialUsed by : Unitree G1 EDU (Orin NX), Skydio X10, NVIDIA Isaac GR00T reference platforms
Sources : Hackster.io (2022-03-22)NVIDIA (technical brief v1.2) (2022-07-01)Futurology.tech (2026-07-09)Skydio (2026-07-09) Last verified 2026-07-09 |
|||||||
| Qualcomm Dragonwing IQ-9075 (IQ9 Series) | Qualcomm | US | SoC | 100 TOPS dense / 200 TOPS sparse (INT8, Hexagon NPU) | n/a | n/a | EVK and modules available (2025-2026); flagship Dragonwing IQ-10 for humanoids in early access June 2026, global availability targeted September 2026 |
|
Qualcomm's frontal assault on NVIDIA Jetson: a full robotics stack unveiled at CES 2026, from household robots to full-size humanoids, with Figure and KUKA among named partners and a power-efficiency pitch inherited from mobile.
Memory : 36 GB LPDDR5 (EVK configuration)Form factor : SoC; EVK, SMARC modules and robot controllers from Lantronix, Advantech, TriaApplications : humanoid, AMR, industrial, drone, visionUsed by : Figure (partnership announced CES 2026), Booster Robotics (partner), KUKA Robotics (partner)
Sources : Qualcomm (2026-07-09)Qualcomm Newsroom (2026-01-06)Lantronix (2026-07-09)Electronics Media (2026-06-29) Last verified 2026-07-09 |
|||||||
| Hailo-10H | Hailo | IL | accelerator | 40 TOPS INT4 / 20 TOPS INT8 | 2.5 | n/a | GA since July 2025; automotive grade (AEC-Q100) with 2026 start of production |
|
First market-available discrete edge accelerator to run LLMs and VLMs fully on-device at ~2.5 W typical, positioning Israel's Hailo as the low-power counterweight to Jetson for drones and vision robotics.
Memory : On-module DRAM via direct DDR interface (M.2 2242/2280 module)Form factor : Discrete accelerator chip; M.2 Key M moduleApplications : drone, vision, AMR, automotive, edge GenAIUsed by : Raspberry Pi AI HAT+ 2 (dev platform), HP and Fujitsu edge products (integrations)
Sources : Hailo (2025-07-22)SiliconANGLE (2025-07-22)Hailo (2026-07-09)Hackster.io (2024-04-10) Last verified 2026-07-09 |
|||||||
| Axelera AI Metis AIPU | Axelera AI | NL | accelerator | 214 TOPS INT8 (peak, quad-core AIPU; 15 TOPS/W) | 8 | $149 | Shipping (M.2/PCIe via distribution); next-gen Europa AIPU (629 TOPS INT8) shipments from H1 2026 |
|
Europe's best-funded edge-AI chip startup (Eindhoven, NL): digital in-memory computing on RISC-V delivering 214 TOPS for ~$149, with the 629-TOPS Europa AIPU and a EUR 61.6M EuroHPC grant anchoring Europe's inference-sovereignty push.
Memory : 1 GB DRAM (M.2 card); 4-16 GB (PCIe card)Form factor : M.2 2280 M-Key card (3.5-9 W typ.) and PCIe HHHL card (8-15 W typ.)Applications : vision, AMR, industrial, retail analytics
Sources : Axelera AI (2026-07-09)Axelera AI Store (2026-07-09)Business Wire (Axelera) (2025-10-21)CNX Software (2023-01-02) Last verified 2026-07-09 |
|||||||
| Tesla FSD Computer AI4 (HW4) | Tesla | US | in-house chip | Not officially disclosed; third-party estimates ≈100–150 TOPS INT8 for the dual‑SoC AI4 system | n/a | n/a | In production (Tesla vehicles and Optimus); successor AI5 taped out, engineering samples expected late 2026, volume 2027 |
|
The only humanoid brain built fully in-house at automotive scale: the same AI4 computer runs Tesla cars and Optimus, and Musk calls the 8x-faster AI5 'absolutely critical' for Optimus, admitting AI4 is not enough for the humanoid program.
Memory : GDDR6 (~384 GB/s per third-party analyses); AI4+ variant announced 2026 with doubled memory (32 GB per SoC)Form factor : Dual-SoC redundant compute board shared between vehicles and OptimusApplications : humanoid, automotiveUsed by : Tesla Optimus
Sources : Electrek (2025-11-25)Wikipedia (2026-07-09)Not a Tesla App (2026-07-09)Teslarati (2026-07-09) Last verified 2026-07-09 |
|||||||
| SiMa.ai MLSoC Modalix | SiMa.ai | US | SoC | 50 TOPS (machine learning accelerator; multi-modal GenAI + CNN) | 10 | $349 | In production since August 2025 (sampling since January 2025) |
|
One of the first chips marketed explicitly as 'Physical AI' silicon: runs LLMs, VLMs and classic computer vision on one SoC under 10 W, targeting robotics and industrial edge as a low-power alternative to Jetson.
Memory : 8-32 GB LPDDR5 (SoM variants)Form factor : SoC + pin-compatible SoM (8 GB $349 / 32 GB $599); DevKit $1,499Applications : vision, AMR, industrial, defense, healthcare
Sources : SiMa.ai (2025-08-12)Business Wire (SiMa.ai) (2025-01-29)SiMa.ai (product brief) (2025-09-09) Last verified 2026-07-09 |
|||||||
| D-Robotics RDK S100 (Horizon Robotics group) | D-Robotics (Horizon Robotics spin-off) | CN | SoC | 80 TOPS (S100) / 128 TOPS (S100P) BPU (INT8-equivalent) | n/a | n/a | GA in China since 2025; international push at Embedded World 2026 (street price ~$389-549 via distributors) |
|
China's main domestic alternative to Jetson for robot builders: the first single-SoC kit merging AI perception and real-time motion control, already adopted by 20+ embodied-AI companies including Leju and LimX Dynamics.
Memory : 12 GB LPDDR5 (96-bit)Form factor : Single-SoC computation-control board: 6x Cortex-A78AE + BPU + real-time MCU cores for motion controlApplications : humanoid, quadruped, AMR, manipulation, visionUsed by : Unitree G1 (launch demo), Leju Robotics humanoids, LimX Dynamics
Sources : D-Robotics (2026-07-09)Pistiz (2026-07-09)Electromaker (2026-07-09)Waveshare (2026-07-09) Last verified 2026-07-09 |
|||||||
| Kalray K300 (MPPA3-80 Coolidge2) | Kalray | FR | accelerator | 40 TOPS INT8 / 20 TFLOPS FP16 (K300 card, 80-core MPPA3-80 v2) | n/a | n/a | Shipping; company refocused on semiconductors in 2025 after divesting its data-acceleration platform to DataCore |
|
France's manycore pioneer survived a 2025 liquidity crisis by refocusing on its MPPA/DPU silicon (semiconductor revenue nearly quadrupled to EUR 14.3M); with GrAI Matter Labs absorbed by Snap, Kalray is one of the last French embedded-AI processor players.
Memory : 16 GB DDR4 @ 3200 MT/sForm factor : PCIe Gen4 card with 4x 10/25G SFP28; 80-core VLIW manycore DPUApplications : vision, industrial, defense, edge AI
Sources : Kalray (2026-07-09)Infonet (2026-07-09)Le Journal des Entreprises (2026-07-09) Last verified 2026-07-09 |
|||||||
| Rockchip RK3588 | Rockchip | CN | SoC | 6 TOPS (3-core NPU; ~2 TOPS/core; supports INT4/INT8/INT16/BF16/TF32; optimized for INT8) | n/a | n/a | GA since 2022; massive installed base across robot controllers and SBCs |
|
The budget workhorse of physical AI: its modest 6-TOPS NPU plus strong CPU/GPU make it the default brain of countless AMRs, robot dev boards and Chinese robotics products where a Jetson would be overkill.
Memory : External LPDDR4/4x/LPDDR5 (board-dependent)Form factor : SoC (8x Cortex-A76/A55 + Mali-G610); ubiquitous on SBCs, SMARC and robot control boards; -40 to +85 C industrial variantsApplications : AMR, vision, education robots, industrialUsed by : Theobroma Systems JAGUAR mobile-robot computer, Dusun IoT AGV/AMR controllers
Sources : TinyComputers.io (2026-07-09)Dusun IoT (2026-07-09)Robotics 24/7 (2026-07-09) Last verified 2026-07-09 |
|||||||
10 / 10 entries