Ahead of its GTC 2024 conference, Nvidia teased that Jensen Huang would reveal groundbreaking advances in accelerated computing, generative AI, and robotics. In the midst of AI's meteoric rise, Nvidia's GTC 2024 announcements were nothing short of a blockbuster event. And as promised, Huang delivered his bombshell this morning.
Nvidia Unveils AI Chip with 25x Cost-Energy Improvement
In his "Accelerated Computing Unveiled" keynote, Nvidia announced the Blackwell, its next-generation AI graphics processing unit (GPU), calling it "very, very powerful." Based on the Blackwell architecture, Nvidia will offer the B200 and GB200 series of chips.
The Blackwell platform reportedly enables building and running real-time generative AI on trillion-parameter large language models (LLMs) at 25x lower cost and energy consumption than its predecessor. Nvidia claims the Blackwell architecture represents the most powerful family of AI chips ever created.
Blackwell 架构 GPU
NVIDIA DGX SuperPOD
NVIDIA H100 Tensor Core GPU compared to the GB200 superchip, delivers up to 30x better performance for large language model inference workloads.NVIDIA GB200 also brings a significant boost in performance. Each DGX GB200 system reportedly packs 36 NVIDIA GB200 Superchips, which themselves feature 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell GPUs. The superchips are connected together by the fifth-generation NVIDIA NVLink to act as a single giant computing engine.
NVIDIA DGX SuperPODs powered by DGX GB200 and DGX B200 systems will be available later this year. “NVIDIA DGX AI supercomputers are the factories for advancing the AI revolution,” said Jensen Huang. “The next generation of DGX SuperPODs combine all of NVIDIA’s latest advances in accelerated computing, networking and software to help every enterprise, industry and nation perfect and deploy their own AI.”
NVIDIA Launches Suite of Microservices
During the keynote, Huang also announced the launch of AI microservices for building and deploying custom applications on their platforms. “The future of software development may be about putting together a bunch of NIMs (NVIDIA inference microservices) to get something trained, deployed,” said Huang, as NVIDIA positions itself as a “foundry” for AI software.
The catalog of cloud-native microservices are built on NVIDIA’s CUDA platform and include NVIDIA’s NIM (NVIDIA Inference Microservice), which is optimized for inference with over 20 of the most popular AI models from NVIDIA and their ecosystem of partners. In terms of performance, NIM delivers pre-built containers based on NVIDIA’s inference software, including the Triton Inference Server and TensorRT-LLM, enabling developers to reduce deployment time from weeks down to minutes.
NVIDIA also announced the availability of 20+ healthcare NVIDIA Medical Imaging and CUDA-X microservices. "These curated microservices add another layer to NVIDIA's full-stack computing platform, bridging the AI ecosystem of model creators, platform providers, and enterprises so they can run customized AI models optimized for the NVIDIA CUDA installed base—billions of GPUs in the cloud, data center, workstations, and PCs," said Jensen Huang.
The NVIDIA ecosystem also includes data, infrastructure, and compute platform providers that are using NVIDIA microservices to bring generative AI to enterprises, along with leading application providers.
Top data platform providers, including Box, Cloudera, Cohesity, Datastax, Dropbox, and NetApp are using NVIDIA microservices to help customers optimize RAG pipelines and integrate proprietary data into generative AI applications.
The topic of humanoid robotics was also a highlight of Huang's keynote. "We can expect humanoids to play a large role in our world. The way that we set up workstations, manufacturing, and logistics, was not designed for humans. And so, these can be deployed more effectively," he said.
Huang unveiled Project GR00T, a Universal Foundation Model for Humanoid Robots, and announced the Jetson Thor, a new humanoid edge computer powered by the NVIDIA Thor system-on-a-chip (SoC), and major updates to the NVIDIA Isaac robotics platform during his keynote.
NVIDIA's new Isaac tools, including Isaac Lab, are designed to create foundational models for robots in any environment. Isaac Lab is a GPU-accelerated, lightweight application optimized for running thousands of parallel simulations for robotics training. OSMO coordinates data generation, model training, and hardware-in-the-loop workflows in a distributed environment.
NVIDIA's latest Jetson Thor platform can perform complex tasks, interact naturally with humans and machines, and has a modular architecture optimized for performance, power efficiency, and size.
NVIDIA is also developing an AI platform for humanoid robots in partnership with 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary.AI, Unitree Robotics, and Xiaopeng Robotics.
NVIDIA announced that leading companies in transportation are adopting the NVIDIA DRIVE Thor centralized vehicle computing platform, including electric vehicle (EV) makers, truck makers, robotaxi and robo-delivery companies, and autonomous bus manufacturers.
DRIVE Thor integrates cockpit functionality, safe and secure highly automated and autonomous driving capabilities onto a single, centralized platform. The next-generation AV platform features the NVIDIA Blackwell architecture, designed specifically for transformer, large language model (LLM), and generative AI workloads.
BYD, GAC Aion, XPeng, Li Auto, and Zeekr have announced that they will build their next-generation vehicles on DRIVE Thor. Plus, Waabi, WeRide, and Nuro will use DRIVE Thor for innovation and validation. DRIVE Thor is expected to be in production as early as next year.
NVIDIA also announced the X800 series, a new family of networking switches purpose-built for massively scaled AI. The NVIDIA Quantum-X800 InfiniBand network and the NVIDIA Spectrum-X800 Ethernet network are the world's first 800Gb/s end-to-end throughput networking platforms, taking network performance for compute and AI workloads to new heights.
