NVIDIA announces the personal AI supercomputer 'Project Digits' - condensing the processing speed of 1 petaflops into desktop size.

At CES 2025 in Las Vegas, NVIDIA announced the personal AI supercomputer 'Project Digits' capable of running AI models with up to 200 billion parameters on the desktop. CEO Jensen Huang stated in the keynote speech that 'AI will become mainstream in all industries and applications,' unveiling a new AI development platform for researchers and students.

The core of Project Digits, the GB10 Grace Blackwell super chip, is a result of a groundbreaking joint development by NVIDIA and MediaTek. This System on Chip (SoC) achieves superior processing performance and power efficiency by connecting the latest...

A feature of the processor configuration includes having a total of 20 CPU cores adopting Arm architecture. Specifically, it employs 10 Arm Cortex-X925 cores for top performance, and 10 Cortex-A725 cores for superior power efficiency...
There are notable features in the memory subsystem, with a total integrated memory of 128GB in the system, enabling coherent memory access between the CPU and GPU. Furthermore, it supports up to 4TB of NVMe storage, enabling fast access to large AI models...

The design of GB10 achieved through collaboration with MediaTek has set an industry benchmark in power efficiency, performance, and connectivity. Particularly through optimization of the power management system...
An AI development environment that can seamlessly integrate with the cloud and Stem Inc.

Project Digits is positioned as NVIDIA's comprehensive solution to the challenge faced by researchers and data scientists in transitioning from development environments to production environments. The underlying software stack of the system adopts the Linux-based NVIDIA DGX OS, and the NVIDIA AI Enterprise software platform is provided with pre-installed.
Developers have access to a rich AI software library and can use various SDKs, orchestration tools, frameworks, and pre-trained models through the NGC (NVIDIA GPU Cloud) catalog and NVIDIA Developer Portal. In particular, models can be fine-tuned using the NVIDIA NeMo framework, and data science processing can be accelerated using the NVIDIA Rapids library. It also seamlessly supports common development environments like PyTorch, Python, and Jupyter Notebook.

Furthermore, one significant feature of Project Digits is its scalability. While it can run AI models with up to 200 billion parameters individually, connecting two Project Digits using NVIDIA ConnectX networking can significantly improve processing power. With the "Double Digits" configuration, models with up to 405 billion parameters can be run, allowing for large-scale models like Meta's Llama-3.1-405B to operate in a local environment. However, super large-scale models like Deepseek-V3 with over 600 billion parameters currently exceed processing capabilities.
For enterprise use cases, NVIDIA AI Enterprise licenses provide enterprise-grade Security, Resistance, and product updates. Additionally, NVIDIA Blueprints and NVIDIA NIM microservices are available for developers of agent-based AI applications for research and development purposes. These tools can be accessed through the NVIDIA Developer Program.
By providing such a comprehensive development environment, developers can experience a consistent development experience from local environment prototyping to production deployment in the cloud or Datacenter. In particular, using the same Grace Blackwell architecture and software platform makes it significantly more efficient to streamline workflows from development to deployment.

Following the announcement of Project Digits, specific expectations and evaluations have been expressed by major players in the AI industry. Particularly noteworthy is the response from companies known for developing large-scale language models.
Ahmad Al-Dahle, the head of the GenAI department at Meta, has commented on the focus of Project Digits enabling local execution of the Llama model, stating that they are working on democratizing access to cutting-edge AI technology through the open-sourcing of the Llama model. With Project Digits, developers can utilize Llama in a local environment, opening up new possibilities for innovation and collaboration.
Silvio Savarese, Chief Scientist at Salesforce, provided an evaluation focusing on the miniaturization and performance aspects of hardware, emphasizing the importance of tools that allow researchers to conduct experiments with scale, speed, and accuracy to advance AI. Project Digits, achieving 128GB of memory in such a small form factor, is said to significantly advance the future of enterprise AI.

Je Boudier, Head of Product at Hugging Face, a rapidly growing AI development platform, shared comments from a more practical perspective. Their goal is to make it easy for developers to build their own AI. With Project Digits, AI builders can now build and run their own GenAI models and systems at the edge. With 128GB of integrated memory, running a 20 billion-parameter model locally and connecting multiple Project Digits systems allows for further scaling, showing optimism for the platform's potential.
Michael Chiang, Co-Founder of Ollama, providing an open-source AI model execution platform, particularly highlighted the scalability of the system. Describing Project Digits as a powerhouse in the palm of your hand, he praised the ease with which developers can handle AI models of up to 40.5 billion parameters by combining two Project Digits.
The evaluations from these industry leaders focus on the feature of Project Digits enabling 'large-scale AI model execution in a desktop environment', emphasizing the importance of development and testing in local environments. This response suggests a renewed recognition of the significance of development and testing in local environments in an era dominated by cloud computing.