Technology Conference The GPU Technology Conference (GTC) 2025, held March 17-21 in San Jose, California, showcased major advances from NVIDIA in artificial intelligence (AI), high-performance hardware, and data center infrastructure. RISC was present at the event, following the innovations that will transform the market.
Here are some of the new features this year:
NVIDIA DGX Spark: Powerful and Scalable AI Mini PC
NVIDIA DGX Spark is a compact and innovative solution for AI development. Equipped with 20 Arm cores connected via C2C to a Blackwell GPU, 128 GB of LPDDR5X memory and 200 GbE NVIDIA ConnectX-7 networking, the device combines power and portability. Its compact design, inspired by the NVIDIA DGX-1, allows for easy portability without compromising efficiency. With only 170 W of consumption, it stands out as an efficient alternative to traditional power-hungry rack servers.
NVIDIA Introduces NVL144 (NVL576) Rack with Rubin Ultra GPUs
NVIDIA surprised by introducing the NVL144 rack, also called NVL576, instead of the expected NVL288. Contrary to expectations, each rack bay contains 18 compute blades instead of 36, directly impacting the overall configuration. With this new architecture, each rack has 144 Rubin Ultra GPUs (2 × 18 × 4), for a total of 576 GPU dies (4 × 144). This change may indicate a new approach by NVIDIA to optimize performance and scalability in accelerated computing systems.
NVIDIA Blackwell Ultra: The New Era of AI Reasoning
NVIDIA today introduced the Blackwell Ultra AI Factory, a platform designed to power AI reasoning with greater performance and efficiency. With new hardware and software solutions, Blackwell Ultra enables complex workloads such as agent AI and physical AI, while reducing operational costs. Key innovations include: NVIDIA GB300 NVL72, connecting 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs; NVIDIA HGX B300 NVL16, delivering up to 11x faster inference than the previous generation; NVIDIA Spectrum-X, enhancing connectivity with 800 Gb/s of throughput per GPU. Additionally, NVIDIA introduced NVIDIA Dynamo, an inference software that improves the efficiency of AI models and reduces response times. Cloud vendors and providers such as AWS, Google Cloud, Microsoft Azure and HPE are already preparing to offer products based on Blackwell Ultra starting in the second half of 2025.
GTC 2025 Key Highlights
Generative AI and Autonomous Agents – Use of tokens for prediction and data analysis, allowing AI to generate content and make decisions independently.
Blackwell GPUs – 30% more efficient models, including the RTX Pro 6000, with 96 GB of memory and optimization for data centers.
AI Infrastructure – Launch of NVIDIA Dynamo, an operating system for AI data centers, and Spectrum X, an advanced network for communication between GPUs.
Strategic Partnerships – NVIDIA collaborations with GM (self-driving cars), Samsung (advanced chips) and Cisco/T-Mobile (Edge AI).
Advanced Robotics – Introducing the GR00T N1, an open-source humanoid robot, and the Isaac GR00T platform, for developing intelligent robots.
Vera Rubin Superchips – Announcing the next generation of exascale AI chips, with launches scheduled for 2026 and 2027.
RISC, with its expertise in infrastructure and technological innovation, continues to monitor these advances to offer the best AI and data center solutions to its customers. Being at GTC 2025 reinforces our commitment to bringing the future of technology to companies seeking efficiency, scalability and innovation.
