Nvidia unveiled the GR00T, a general-purpose large model in the field of robotics, and the next-generation chip architecture Blackwell

Odaily Planet Daily News Nvidia founder Jensen Huang delivered an opening speech at the GTC conference held in San Jose, USA. In this presentation, NVIDIA unveiled their next-generation chip architecture, Blackwell. According to reports, the Blackwell GPU is named after mathematician David Harold Blackwell, and it also follows the Hopper architecture previously launched by NVIDIA. With 208 billion transistors in a Blackwell GPU, it can support AI models with up to 10 trillion parameters. In addition to the chip itself, this architecture also uses the fifth-generation NVLink high-speed interconnect and the second-generation Transformer engine, which is comprehensively upgraded in many aspects. According to Huang, the new chip will be available later in 2024. Huang said that in the future, Nvidia plans to use Blackwell to expand into AI companies around the world, signing contracts with all OEMs, regional clouds, national sovereign AI, and telecommunications companies around the world. At present, Amazon, Dell, Google, Meta, Microsoft, OpenAI, and Tesla have all planned to use Blackwell GPUs. It is worth noting that the legendary B100 did not appear before, and Nvidia instead released the GB200, a superchip consisting of a Grace CPU and two sets of Blackwell GPUs. According to Huang, the GB200 has six times the computing power of the H100, and the computing power of the corresponding multimodal specific domain can reach 30 times. In addition, Nvidia also announced the GB200 NVL72, a server consisting of 36 Grace CPUs and 72 Blackwell GPUs. This time, Nvidia will launch a general basic large model for the field of robotics, called GR00T. In addition, NVIDIA has introduced a new computer for robotics, Thor, which has been specifically optimized for performance, power consumption, and size. NVIDIA ALSO INTRODUCED A NEW AI INFERENCE SERVER, NIM (NVIDIA INFERENCE MICROSERVICE), WHICH ALLOWS ANYONE TO CUSTOMIZE AI MODELS AND APPLICATIONS IN THIS FORM.