NVIDIA concluded its 16th GPU Technology Conference (GTC) last Thursday. GTC 2024 commemorates two years since CEO Jensen Huang unveiled the Hopper architecture-based GPUs at GTC 2022, which quickly became the preferred option for AI developers in both training and inferencing tasks. Positioned at the forefront of the AI revolution, NVIDIA unveiled a range of new offerings at GTC 2024. Explore these latest innovations further.
It certainly feels like the “Era of AI,” as proclaimed by NVIDIA, given each week, there’s AI galore. Consider this week:
The United Nations General Assembly adopted a global resolution for AI
Apple and Google are discussing leveraging the latter’s Gemini LLM for the former’s devices
Google DeepMind and the Liverpool Football Club collaborated to develop TacticAI, an AI soccer assistant
Stability AI released Stable Video 3D (SV3D) to convert images into videos
Sakana AI developed an evolutionary algorithms-based merging technique for basic model creation
Google penalized €250 million (about $272 million) for using copyrighted material from publishers in its AI tech
xAI open-sources the Grok-1 LLM
Apple CEO revealed the company is working on MM1, a generative AI model with 64 billion parameters
And NVIDIA’s chips power nearly all of this. A testament to this fact is the company’s soaring stock price, which has increased by 89.83% year-to-date and by 245.45% in the previous 12 months, taking its market capitalization to over $2.28 trillion and making it the third-most valued company in the world, and its CEO, the black leather jacket-clad Huang, as one of the most important figures in tech.
And NVIDIA isn’t planning to slow down anytime soon.
Moreover, it’s not just about the hardware. NVIDIA’s offerings are steadily enabling the company to morph into a full-stack AI development engine. Take a look at what NVIDIA had to show at GTC 2024.
NVIDIA GTC 2024 Highlights 1. Blackwell GPUs
Huang took the stage to introduce raw compute gains for AI dev through the new Blackwell GPUs. The flagship new GB200 chip is designed using two Blackwell NVIDIA B200 Tensor Core GPUs with a new Transformer Engine (the company’s way of saying 4-bit floating point or FP4) and a new GB200NVL 72 cluster capable of delivering 30% higher performance.
Blackwell chips will be available later this year and have multiple takers, including Microsoft, Amazon, etc.
- Scalable AI infrastructure
Keeping in mind the future demands, NVIDIA announced the DGX SuperPOD. It consists of multiple DGX GB200 systems, each comprising 36 NVIDIA GB200 Superchips (36 Grace Hopper CPUs and 72 Blackwell GPUs) joined by the new fifth-gen NVLink to deliver 11.5 exaflops of performance at FP4 precision and 240 terabytes of memory.
Based on the rack-scale architecture, this can be scaled up to eight or more DGX GB200 systems with tens of thousands of GB200 Superchips connected via NVIDIA Quantum InfiniBand. A monster indeed.
Amazon Web Services has reportedly upgraded its Project Ceiba to include the new Blackwell GPUs now scaled through the NVIDIA GB200 NVL72. Originally planned to deliver a mammoth 65 exaflops, the machine developed under Project Ceiba will now offer 414 exaflops, thanks to AI scaling capacity with 20,736 GB200 Superchips.
- NVIDIA-MediaTek partnership
As expected, NVIDIA licensed its GPU tech to MediaTek under a new collaboration that saw the latter develop and introduce advanced system-on-chips (SoCs) for the automotive industry. The four new chips, CX-1, CY-1, CM-1, and CV-1, are a part of the Dimensity Auto Cockpit MediaTek unveiled at GTC 2024.
The partnership intends to bring AI processing and experiences directly to vehicles instead of needing a cloud connection. They are based on the NVIDIA RTX GPU for realistic ray tracing and feature Arm V9-A CPU cores, multi-camera support, multi-audio DSP, and an HDR image signal processor (ISP).
- Project GR00T
Project GR00T, or Generalist Robot 00 Technology, is NVIDIA’s “general-purpose foundation model for humanoid robots.” Humanoid robots have been a topic of discussion in science fiction for decades, but with AI catching up, NVIDIA seems keen not to miss out.
The project includes Jetson Thor, a computer designed to run multimodal generative AI models like Gr00T, and is based on Blackwell GPUs. “Building foundation models for general humanoid robots is one of the most exciting problems to solve in AI today,” Huang said. “The enabling technologies are coming together for leading roboticists around the world to take giant leaps towards artificial general robotics.”
From what was revealed, NVIDIA is in the initial stages of developing a robot model capable of emulating humans and performing human-centric tasks. While it does, the company said it is creating an AI platform for humanoid robot companies 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree Robotics and XPENG Robotics, and others.
- API availability
To usher in greater adoption of its metaverse simulation offering, the Omniverse Cloud, NVIDIA is making it available through application programming interfaces (APIs). Used for design, simulation, and other operational elements by designers, engineers, and software developers, API-based availability opens the 3D modeling tech to a wide range of industrial digital twin applications.
NVIDIA Omniverse Cloud will be available through five APIs (USD Render, USD Write, USD Query, USD Notify, and Omniverse Channel). It can be integrated into existing design and automation software applications.
“Through the NVIDIA Omniverse API, Siemens empowers customers with generative AI to make their physics-based digital twins even more immersive,” said Roland Busch, president and CEO of Siemens AG. “This will help everybody to design, build, and test next-generation products, manufacturing processes, and factories virtually before they are built in the physical world. By combining the real and the digital worlds, Siemens digital twin technology is enabling companies around the world to become more competitive, resilient, and sustainable.”
Siemens is one of the companies that has backed Omniverse Cloud APIs, the others being Microsoft, Dassault Systèmes, Rockwell Automation, Hexagon, Cadence, Trimble, and Ansys.
The APIs also mean that users can stream Omniverse OpenUSD files onto the Apple Vision Pro, making the spatial computing device attractive for enterprise and industrial use cases.
- Bridging the AI divide
NVIDIA hopes to bring AI development and deployment together with enterprise operations by reducing the time it takes to deploy inference software with NVIDIA Inference Microservices (NIM). These cloud-native microservices for generative AI provide access to pre-built containers, models, APIs, and inference engines on the NVIDIA Enterprise AI platform.
“Established enterprise platforms are sitting on a goldmine of data that can be transformed into generative AI copilots,” Huang said. “Created with our partner ecosystem, these containerized AI microservices are the building blocks for enterprises in every industry to become AI companies.”
- Partner launches
At SXSW, Dell founder and CEO Michael Dell estimated data center demand to surge by over 100x in the next decade, thanks to the AI revolution. A week later, at GTC 2024, the company announced the Dell AI Factory, which integrates Dell’s computing, storage, client device, software, and services capabilities with NVIDIA’s advanced AI infrastructure and software suite for custom AI model training.
Dell catering for future demands isn’t lost on others. Hewlett Packard Enterprise, Supermicro, and Lenovo also announced respective generative AI infrastructure in collaboration with NVIDIA at GTC 2024.
ClearML introduced Fractional GPU, a new tool to manage NVIDIA GPUs. It relies on two features NVIDIA ships with its GPUs—multi-instance GPU and time slicing—to optimize GPU utilization. The tool enables developers to cater to multiple AI workloads on a single GPU.
Balbix’s BX4 AI Engine cyber risk management tool is based on NVIDIA’s tech, including the full-stack NVIDIA AI platform and Nvidia GPUs. It is also integrated with the NVIDIA Triton Inference Server, NVIDIA TensorRT-LLM. The BX4 AI Engine offers asset and vulnerability visibility, risk reduction, and security control measures.
On the storage side, Pure Storage introduced the new validated reference architecture for generative AI, which the company developed in collaboration with NVIDIA. Based on the Retrieval-Augmented Generation (RAG) Pipeline for AI Inference, it caters to AI-specific data management and computing requirements.
Source : Spiceworks
