NVIDIA GTC, the corporate’s annual AI convention, simply kicked off with CEO Jensen Huang taking the stage to disclose new updates from the corporate.
One of many major bulletins that Huang made was the Blackwell platform, which is a GPU structure for working generative AI. In keeping with the corporate, it may well run “real-time generative AI on trillion-parameter giant language fashions at as much as 25x much less value and power consumption than its predecessor,” Hopper.
Blackwell consists of six applied sciences: a 208 billion transistor chip, a second-gen transformer engine, fifth-generation NLink, RAS engine, decompression engine, and assist for brand new native interface encryption protocols.
The corporate believes that Blackwell will allow breakthroughs in areas like information processing, engineering simulation, digital design automation, computer-aided drug design, quantum computing and generative AI.
Various firms are anticipated to undertake Blackwell, together with AWS, Dell, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and xAI.
“For 3 a long time we’ve pursued accelerated computing, with the objective of enabling transformative breakthroughs like deep studying and AI,” stated Huang. “Generative AI is the defining expertise of our time. Blackwell is the engine to energy this new industrial revolution. Working with essentially the most dynamic firms on the earth, we are going to notice the promise of AI for each trade.”
The corporate additionally introduced a brand new supercomputer, DGX SuperPOD, that’s powered by the Blackwell chips. The techniques embody 36 GB200 Superchips (which every comprise 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell GPUs). In keeping with NVIDIA, these Superchips every ship a 30x efficiency enhance for LLM inference in comparison with NVIDIA H100 Tensor Core GPUs.
It may additionally scale as much as tens of hundreds of Superchips utilizing NVIDIA’s Quantum InfiniBand.
SuperPOD additionally gives predictive-management capabilities to assist lower down on downtime and inefficiency, and in addition identifies areas of concern in order that it may well recommend upkeep steps, modify compute sources, and save and resume jobs.
“NVIDIA DGX AI supercomputers are the factories of the AI industrial revolution,” stated Huang. “The brand new DGX SuperPOD combines the newest developments in NVIDIA accelerated computing, networking and software program to allow each firm, trade and nation to refine and generate their very own AI.”
