ZED Box 2 Max

The ZED Box 2 Max is an industrial edge AI computer engineered specifically for physical AI, multi-camera spatial perception, and autonomous robotics. Powered by the NVIDIA Jetson Thor system-on-chip, the platform integrates Blackwell architecture compute with unified memory architectures to run high-throughput vision-language-action (VLA) models, deep learning perception stacks, and large language models at the physical edge. Originally developed by Stereolabs (an Ouster company) and cataloged under Unknown hardware for AI development, the unit addresses a critical gap in edge hardware: pairing enterprise-grade sensor ingestion with raw generative AI throughput.

Priced from $5,999 for the base Jetson T4000 configuration, the ZED Box 2 Max sits in the industrial edge and prosumer robotics tier. It does not compete with standard desktop workstations or server-rack units intended for climate-controlled data centers. Instead, it is built for deployment on mobile robots, industrial autonomous machines, humanoids, and smart infrastructure where high-vibration resilience, wide voltage input (9 to 36 V DC), and deterministic multi-sensor synchronization are mandatory.

For teams building autonomous physical agents, this system serves as an all-in-one compute and perception engine. By combining up to eight hardware-synchronized GMSL2 camera links with up to 128 GB of unified memory, the ZED Box 2 Max handles concurrent sensor fusion, localized SLAM, and dense transformer model inference in real time without offloading frames to remote cloud infrastructure.

When evaluating the ZED Box 2 Max for AI workloads, the focal point is the NVIDIA Jetson Thor module (available in T4000 and T5000 variants). The flagship T5000 configuration delivers up to 2,070 FP4 TFLOPS (sparse) of compute via a 2,560-core Blackwell GPU equipped with fifth-generation Tensor Cores. The T4000 tier provides 1,200 FP4 TFLOPS across 1,536 GPU cores. On the host side, the platform relies on a 12-core or 14-core Arm Neoverse-V3AE CPU clocked up to 2.6 GHz, delivering significant multi-threaded overhead for processing point clouds, path planning, and ROS node graphs.

The memory subsystem is configured with up to 128 GB of 256-bit LPDDR5X unified memory offering 273 GB/s of bandwidth. Because the CPU, GPU, and vision accelerators share a single pool, there is no PCIe transfer penalty when passing high-resolution camera feeds directly into deep neural network inference pipelines.

While 128 GB of capacity places it in the upper echelon of edge devices, the 273 GB/s memory bandwidth is the primary operational metric to consider for autoregressive language generation. Modern desktop workstations with dual consumer GPUs can exceed 1,000 GB/s, meaning pure token streaming for text-heavy LLMs will be constrained compared to desktop hardware. However, for edge devices, 273 GB/s provides a balanced envelope between high memory capacity and mobile battery conservation.

• AI Compute: Up to 2,070 FP4 TFLOPS (sparse) on Jetson T5000; 1,200 FP4 TFLOPS on Jetson T4000
• SoC Architecture: NVIDIA Jetson Thor (Blackwell GPU + Arm Neoverse-V3AE CPU)
• Unified Memory: 64 GB (T4000) or 128 GB (T5000) LPDDR5X at 273 GB/s
• Camera Ingestion: 8x GMSL2 inputs via dual Quad Fakra-Z connectors with hardware frame synchronization
• Networking: 1x QSFP supporting 4x 10 GbE, plus 2x Gigabit Ethernet
• Storage: Dual NVMe SSD expansion slots (256 GB, 1 TB, or 2 TB configurations)
• Peripheral I/O: 4x CAN, GPIO, 2x USB, HDMI, Dual-RTK GNSS support, and optional 4x PoE+
• Input Power: 9 to 36 V DC with integrated ESD, transient, and reverse-polarity protection
• Configurable Module Power Modes: 70W, 90W, 120W, and MAXN (entire system max power up to 250W)

Having a 128GB GPU for AI in an embedded form factor fundamentally changes edge model deployment, eliminating the standard 16 GB or 32 GB memory ceiling typical of small-footprint systems.

Because unified memory is shared between system processes and the GPU, the ZED Box 2 Max VRAM for large language models comfortably supports large parameter architectures without CPU offloading:

• 70B to 72B Parameter Models: As hardware for running 70B parameter models at the edge, the 128 GB memory ceiling easily accommodates Llama 3.1 70B or Qwen 2.5 72B quantized to INT4 or FP4 (taking roughly 40 to 45 GB of footprint), leaving tens of gigabytes free for dynamic context caching and operating system overhead.
• Medium Models (14B to 32B): Architectures such as Qwen 2.5 14B/32B, Mistral 8x7B, or specialized reasoning distillations like DeepSeek-R1-Distill-Qwen-32B run natively at FP8 or INT8 with full 32k to 128k context windows.
• Small Frontier Models (7B to 8B): Llama 3.1 8B and Mistral 7B run completely unquantized at FP16 or BF16.

Memory bandwidth directly bounds ZED Box 2 Max tokens per second during autoregressive decoding. Based on the 273 GB/s LPDDR5X bus:

• 70B Parameters (INT4 / FP4): Expect around 4 to 6 tokens per second. While not suitable for rapid bulk summarization, this rate is fast enough for tactical decision-making loops in high-level robotic task planning.
• 14B to 32B Parameters (INT4 / FP8): Inference settles between 8 and 18 tokens per second, making this size the sweet spot for balance between reasoning depth and execution speed.
• 7B to 8B Parameters (INT4 / FP8): Delivers 25 to 35+ tokens per second, enabling interactive, conversational robot-human interaction.

The hardware architecture shines brightest on non-autoregressive or hybrid perception models. Running native FP4 Tensor Core acceleration allows models like OpenVLA, Florence-2, and CLIP-based embeddings to execute at near-instantaneous latency across multi-camera streams. The eight GMSL2 camera inputs feed into the hardware video decoders, allowing spatial models to ingest 360-degree vision without stalling the primary model forward passes.

The ZED Box 2 Max is not a budget desktop appliance for hobbyists running local chatbots. It is an industrial-grade compute platform engineered for production deployments.

• Humanoid and Mobile Robotics: Engineers building Autonomous Mobile Robots (AMRs) that require onboard sensor fusion, simultaneous 360-degree SLAM, and local VLA reasoning without network tethering.
• Edge Deployment Scenarios: Defense, agricultural autonomy, and mining environments requiring a ruggedized enclosure that can withstand dirty electrical buses (9-36 V DC) and high shock environments.
• Multi-Sensor AI Agents: Teams using ROS 2 or NVIDIA Isaac ROS to process dense spatial data from 8 hardware-synchronized cameras alongside LiDAR over high-speed 10 GbE links.
• Local AI Agents 2026: Developers creating autonomous agents that need to operate in field-deployed environments with zero latency dependency on cloud APIs.

The system is strictly an edge inference and continuous-learning execution device. While lightweight fine-tuning (such as low-rank adaptation via LoRA) for smaller 7B-class models is feasible given the 128 GB memory pool, prolonged pre-training on this device is impractical. Workflows should center on running pre-trained, heavily quantized checkpoints optimized through TensorRT or Isaac Lab.

The prior standard for rugged robotics edge compute was the Jetson AGX Orin 64GB Industrial module. The ZED Box 2 Max provides a generational jump:

• Compute Throughput: Thor provides up to 2,070 FP4 TFLOPS versus the AGX Orin's 275 INT8 TOPS. The introduction of Blackwell's FP4 Tensor Cores allows users to run models twice as large at lower latency.
• Memory Overhead: Moving from 64 GB on the highest-tier Orin to 128 GB on the Thor T5000 doubles the memory envelope, converting what used to be a system limited to 8B-14B models into a platform capable of hosting full 70B parameter models alongside vision stacks.
• Camera Ingestion: The native integration of eight GMSL2 Fakra inputs saves thousands of dollars in external deserializer carrier boards and integration labor required by generic Orin development kits.

System integrators often evaluate dual desktop GPU configurations (such as custom ruggedized PCs with two RTX 4090 cards) for mobile robotic compute:

• Bandwidth vs. Efficiency: Dual RTX 4090 setups offer over 2,000 GB/s combined memory bandwidth, yielding vastly superior token generation speeds for text LLMs. However, they consume between 600W and 900W of power, generate massive heat, require bulky liquid or active thermal designs, and lack native GMSL2 deserialization.
• Power Envelope: The ZED Box 2 Max caps total system power consumption at 250W, making it viable for battery-powered platforms where thermal dissipation and weight are hard constraints.
• VRAM Limitations: Dual RTX 4090s provide only 48 GB of total VRAM across split memory pools. The ZED Box 2 Max offers 128 GB of unified memory, allowing developers to load massive models without complex pipeline parallelism across PCIe switches.

For teams deploying physical AI agents, humanoid control loops, and multi-camera perception pipelines in field conditions, the ZED Box 2 Max stands out as the best AI chip for local deployment where industrial I/O, ruggedized power, and massive unified memory capacity are absolute operational requirements.