MegaCube-H3C

LinSeer MegaCube equipped with the NVIDIA GB10 Grace Blackwell Superchip, it delivers 1 petaFLOP of computing power in a palm-sized form factor,thoroughly breaking spatial constraints for AI developer desktops. Featuring 128GB of LPDDR5x unified memory, it condenses the capability for single-machine inference of 200B parameter models and fine-tuning of 70B parameter models into a desktop environment. Connecting two units using NVIDIA ConnectX networking enables support for t cutting-edge large models like Llama 3.1 405B. Pre-installed with the NVIDIA AI software stack, it provides research institutions and enterprise teams with an "out-of-the-box" localized large model productivity tool, truly enabling technological accessibility for AI developer communities.

1 PetaFLOP FP4 Compute Power

Natively supports FP4 precision, unleashing 1 petaFLOP of FP4 compute power. The inference speed is significantly enhanced compared to software-simulated FP4 efficiency, delivering ultimate performance.
128GB LPDDR5x
Unified Memory

A single LinSeer MegaCube can support
inference for 200B parameter models.

A single LinSeer MegaCube can support inference for 200B parameter models.
Scalable Performance

High-performance NVIDIA Connect-X networking
allows two LinSeer MegaCube systems to be
interconnected and is compatible with AI models of
up to 405B parameters.

High-performance NVIDIA Connect-X networking allows two LinSeer MegaCube systems to be interconnected and is compatible with AI models of up to 405B parameters.
Ultra-Compact

About the size of a palm. Effortless deployment in labs and offices without space constraints.
Out-of-the-Box Full-Stack AI Ecosystem

Pre-installed with NVIDIA DGX OS, CUDA, CUDA-X, RTX toolkit and
libraries. Compatible with PyTorch, TensorFlow, MATLAB, and more.

Pre-installed with NVIDIA DGX OS, CUDA, CUDA-X, RTX toolkit and libraries. Compatible with PyTorch, TensorFlow, MATLAB, and more.

Compact Size, Immense Capability

In an era where computing power equals productivity, LinSeer MegaCube deeply integrates top-tier hardware with intelligent scheduling, providing a solid foundation for complex AI scenarios:

Great Things Come in Small Packages

Compact Size
150 x 150 x 51mm

Equipped with NVIDIA GB10 Grace
Blackwell Superchip, Delivering
1 PetaFLOP FP4 Compute Power

Equipped with NVIDIA GB10 Grace Blackwell Superchip, Delivering 1 PetaFLOP FP4 Compute Power

NVIDIA Blackwell architecture GPU within the NVIDIA GB10 Grace Blackwell Superchip contains 6144 CUDA cores, equipped with fifth-generation Tensor Cores, unleashing 1 petaFLOP of FP4 compute power. The GB10 also includes a high-performance 20-core ARM CPU, providing robust data preprocessing and scheduling performance to accelerate model inference and fine-tuning.

NVIDIA NVLink™-C2C Technology

NVIDIA NVLink™-C2C breaks down communication barriers between the CPU and GPU, offering bandwidth up to 5 times that of PCIe 5.0. It provides ultra-high bandwidth, ultra-low latency interconnection, allowing the 128GB system unified memory to truly function as a "large model container," significantly enhancing AI computing efficiency.

128GB LPDDR5x
Unified Memory

128GB LPDDR5x Unified Memory

Provides ample memory for efficient AI model inference, fine-tuning, and other tasks. A single LinSeer MegaCube can support inference for 200B parameter models.

NVIDIA ConnectX-7 2*200G
Network Card

The built-in NVIDIA ConnectX-7 technology, allows two LinSeer MegaCube units to be cascaded, thereby supporting inference for models with up to 405B parameters, such as Llama 3.1 405B.

Out-of-the-Box Full-Stack AI Ecosystem

Pre-installed OS and
Development Environment

The LinSeer MegaCube comes pre-installed with NVIDIA DGX OS (Ubuntu-based), optimized for AI development environments, eliminating complex configuration needs.
Pre-installed with the NVIDIA AI software stack, developers can work on projects on their LinSeer MegaCuge right away.
Supports Full-Stack AI Software

LinSeer MegaCube supports AI framework software like PyTorch, TensorFlow, Matlab, AI development platforms like NVIDIA Riva, NVIDIA Holoscan, NVIDIA Metropolis, NVIDIA Isaac as well as access to AI models tuned for LinSeer MegaCube systems with NVIDIA NIM. (Note: "Blueprint" might be a specific term. "NVIDIA AI Enterprise" is common for suite, verify if "Blueprint" is correct).
Mainstream Model
Support

Supports fine-tuning and inference for the Deepseek R1 70B model.
Supports inference for large models like Llama 3.1 405B, Qwen3 235B, etc. (requires cascading two LinSeer MegaCube systems).

Rich Applications Scenarios

Prototyping

Develop, test, and validate AI models and applications.

Leveraging the full NVIDIA AI software stack, it provides developers with a one-stop platform for AI model creation, testing, and validation, enabling easy development of AI-enhanced applications and industry solutions. After local debugging, migrate work to NVIDIA DGX Cloud on leading cloud providers, accelerated data centers, or cloud infrastructure to scale for additional work or deployment.
Fine-Tuning

Customize and optimize the performance of pre-trained models.

Through 128GB of massive unified system memory, it supports fine-tuning for models with up to 70 billion parameters, precisely adapting to vertical industry needs and specific use cases. Whether infusing knowledge for specialized scenarios like healthcare and finance, or adapting models for low-resource languages or niche industry data, it efficiently enhances the task accuracy and practicality of pre-trained model.
Inference

Supports efficient inference validation for 200 billion parameter models.

Easily conduct testing, validation, and deployment of 200 billion parameter models. The hardware-native NVFP4 technology reduces memory footprint while maximizing performance while minimizing accuracy loss.
Data Science

High-performance data science at your desk.

NVIDIA DGX Spark’s combination of 128GB of unified memory and 1 petaFLOP of parallel throughput maximizes performance of large, computationally complex data analytics and machine learning workflows at your desk. Clean your data.
Edge Applications

Develop edge intelligence applications based on NVIDIA professional frameworks.

Provides an excellent development platform for edge scenarios like robotics, smart cities, and computer vision, deeply compatible with frameworks such as NVIDIA Isaac (robotics), Metropolis (intelligent video analytics), and Holoscan (medical imaging).

Cloud & AI

Intelligent Connection

Intelligent Computing

Intelligent Storage

Security

SMB Products

Intelligent Terminal Products

Industry Solutions

Product Support Services

Technical Service Solutions

Resource Center

Policy

Online Help

Technical Blogs

Training & Certification

Become a Partner

Partner Policy & Program

Global Learning

Partner Sales Resources

Partner Business Management

Service Business

Profile

News & Events

Online Exhibition Center

Contact Us