NVIDIA’s Announcements & Future Plans from GTC 2025
25 March 2025 · Uncategorized ·
Source: · https://technews.tw/2025/03/19/nvidia-gtc-2025-jensen-huang-keynote/
At this year’s GPU Technology Conference (GTC), CEO Jensen Huang announced updates across NVIDIA's product lines, including Blackwell Ultra, the next-generation Rubin architecture, and silicon photonics network switch technology. TechNews has summarized key points from his keynote speech to provide a quick overview for readers.
Blackwell Ultra is scheduled for release later this year and features an increased memory capacity of up to 288GB, enabling support for larger AI models.
Huang stated that "AI development is advancing rapidly; inference and agent-based AIs require computational power several orders of magnitude higher than what we currently have. Blackwell Ultra was designed specifically for such a moment—this versatile single platform can efficiently execute pre-training, post-training, and inferential reasoning."
Vera Rubin represents the next generation NVIDIA architecture slated to launch in late 2026. The latest Vera chip and server will offer faster speeds than its predecessor, particularly regarding data transmission between chips. NVIDIA plans on designing their own processor called Vera CPU as a replacement for Grace CPUs currently used.
Additionally, Vera Ultra is scheduled for release by the end of 2027 while Feynman (named after physicist Richard Feynman) will be introduced in late 2028.
NVIDIA recently unveiled DGX Spark and supercomputer DGX Station to enable AI developers, researchers, data scientists, and students to prototype, fine-tune, or infer large models on desktop computers. Users can run these models locally or deploy them onto NVIDIA’s cloud infrastructure.
The new DGX AI computer powered by Blackwell Ultra brings the power of the Grace-Blackwell architecture—previously confined within server rooms—down into a personal workstation that challenges Apple's high-end Macs in performance.
NVIDIA has also integrated optical innovations with its silicon photonics network switch, reducing laser count fourfold while improving energy efficiency thirty-five times and signal integrity sixty-three times. Quantum-X InfiniBand switches will be available by the end of 2025, whereas Spectrum-X Photonics Ethernet switches are slated for release in late 2026.
The platform utilizes TSMC’s COUPE silicon photonics technology, integrating a 65nm electronic integrated circuit (EIC) with photonic ICs using SoIC-X packaging. NVIDIA has partnered with Taiwanese companies such as Foxconn and SPIL alongside American-Chinese-Japanese firms like Coherent Inc., Corning Incorporated to establish an ecosystem for the stable supply of silicon photonics, further solidifying its position in proprietary hardware.
NVIDIA introduced Dynamo, open-source inference software designed to accelerate AI model expansion at minimal cost. The goal is to speed up reasoning processes where models think through problems step-by-step rather than providing immediate answers.
The company also unveiled Isaac GR00T N1 as the world's first open-source humanoid robot foundation model, offering general-purpose robotic capabilities and skills.
Next-generation wireless networks must integrate AI from their foundations to seamlessly connect billions of phones, sensors, cameras, robots, and self-driving cars. An AI-native wireless network will provide stronger services for millions more users while setting new standards for spectral efficiency.
To this end, NVIDIA is collaborating with industry leaders such as T-Mobile USA Inc., MITRE Corporation, Cisco Systems Inc., ODC (a portfolio company of Cerberus Capital Management), and Booz Allen Hamilton to research and develop AI-native 6G wireless network hardware, software, and architecture.
NVIDIA will partner with General Motors Co. using NVIDIA’s accelerated computing platform to create customized AI systems—including the use of Omniverse alongside NVIDIA Cosmos for training manufacturing models that optimize factory planning and robotics technology at GM.
General Motors also plans on utilizing DRIVE AGX onboard hardware in future advanced driver assistance systems as well as enhanced driving experiences inside vehicles.
TechNews 科技新報（@technewsinside）分享的貼文
(Lead image source: NVIDIA GTC)
Blackwell Ultra is scheduled for release later this year and features an increased memory capacity of up to 288GB, enabling support for larger AI models.
Huang stated that "AI development is advancing rapidly; inference and agent-based AIs require computational power several orders of magnitude higher than what we currently have. Blackwell Ultra was designed specifically for such a moment—this versatile single platform can efficiently execute pre-training, post-training, and inferential reasoning."
Vera Rubin represents the next generation NVIDIA architecture slated to launch in late 2026. The latest Vera chip and server will offer faster speeds than its predecessor, particularly regarding data transmission between chips. NVIDIA plans on designing their own processor called Vera CPU as a replacement for Grace CPUs currently used.
Additionally, Vera Ultra is scheduled for release by the end of 2027 while Feynman (named after physicist Richard Feynman) will be introduced in late 2028.
NVIDIA recently unveiled DGX Spark and supercomputer DGX Station to enable AI developers, researchers, data scientists, and students to prototype, fine-tune, or infer large models on desktop computers. Users can run these models locally or deploy them onto NVIDIA’s cloud infrastructure.
The new DGX AI computer powered by Blackwell Ultra brings the power of the Grace-Blackwell architecture—previously confined within server rooms—down into a personal workstation that challenges Apple's high-end Macs in performance.
NVIDIA has also integrated optical innovations with its silicon photonics network switch, reducing laser count fourfold while improving energy efficiency thirty-five times and signal integrity sixty-three times. Quantum-X InfiniBand switches will be available by the end of 2025, whereas Spectrum-X Photonics Ethernet switches are slated for release in late 2026.
The platform utilizes TSMC’s COUPE silicon photonics technology, integrating a 65nm electronic integrated circuit (EIC) with photonic ICs using SoIC-X packaging. NVIDIA has partnered with Taiwanese companies such as Foxconn and SPIL alongside American-Chinese-Japanese firms like Coherent Inc., Corning Incorporated to establish an ecosystem for the stable supply of silicon photonics, further solidifying its position in proprietary hardware.
NVIDIA introduced Dynamo, open-source inference software designed to accelerate AI model expansion at minimal cost. The goal is to speed up reasoning processes where models think through problems step-by-step rather than providing immediate answers.
The company also unveiled Isaac GR00T N1 as the world's first open-source humanoid robot foundation model, offering general-purpose robotic capabilities and skills.
Next-generation wireless networks must integrate AI from their foundations to seamlessly connect billions of phones, sensors, cameras, robots, and self-driving cars. An AI-native wireless network will provide stronger services for millions more users while setting new standards for spectral efficiency.
To this end, NVIDIA is collaborating with industry leaders such as T-Mobile USA Inc., MITRE Corporation, Cisco Systems Inc., ODC (a portfolio company of Cerberus Capital Management), and Booz Allen Hamilton to research and develop AI-native 6G wireless network hardware, software, and architecture.
NVIDIA will partner with General Motors Co. using NVIDIA’s accelerated computing platform to create customized AI systems—including the use of Omniverse alongside NVIDIA Cosmos for training manufacturing models that optimize factory planning and robotics technology at GM.
General Motors also plans on utilizing DRIVE AGX onboard hardware in future advanced driver assistance systems as well as enhanced driving experiences inside vehicles.
TechNews 科技新報（@technewsinside）分享的貼文
(Lead image source: NVIDIA GTC)