On Monday (March 18) local time, the high-profile Nvidia annual GTC conference kicked off in San Jose, California。

After five years, Nvidia once again offline to hold the annual GTC conference, so the expectations of the conference is very high。Nvidia did not disappoint, introducing the latest results of several companies at once.。

The New Blackwell GenerationGPUArchitecture

At this conference, the most eye-catching is Nvidia's new generation of Blackwell GPU architecture。

Blackwell enables organizations around the world to build and run real-time generative AI on trillion-parameter large language models (LLMs), reducing the cost and energy consumption by up to 25 times compared to the previous generation.。

The Blackwell GPU architecture is powered by six transformative accelerated computing technologies:

● The world's most powerful chip: Blackwell-based GPUs with 208 billion transistors, manufactured in TSMC's 4-nanometer (nm) process, connect GPU dies into a unified GPU with 10TB / s inter-chip interconnection。

● Second-generation Transformer engine: Blackwell will double its computing power and model size with new 4-bit floating-point AI inference capabilities, thanks to new microtensor scaling support and Nvidia's advanced dynamic range management algorithms integrated into the Nvidia TensorRT ™ -LLM and NeMo Megatron frameworks。

● Fifth Generation NVLink: In order to improve the performance of trillions of parametric models and hybrid expert AI models, the fifth generation NVLink provides a breakthrough 1 for each GPU..8TB / s bidirectional throughput, ensuring seamless high-speed communication between up to 576 GPUs, which can meet the needs of today's most complex LLM。

● RAS Engine: Blackwell-based GPUs include a dedicated engine for reliability, availability, and maintainability。In addition, the Blackwell architecture adds chip-level capabilities that leverage AI preventive maintenance to run diagnostics and predict reliability-related issues。This will maximize system uptime and improve the resilience of large-scale AI deployments, allowing them to run continuously for weeks or even months, while reducing operating costs。

● Secure AI: Advanced confidential computing capabilities can protect AI models and customer data without compromising performance, and support new local interface encryption protocols, which are essential for industries with high privacy concerns such as healthcare and financial services。

● Decompression engine: A dedicated decompression engine supports the latest formats, providing extremely powerful data analysis and data science performance by accelerating database queries。Data processing, which requires businesses to spend tens of billions of dollars a year, will increasingly be GPU-accelerated in the coming years.。

With these technologies, Blackwell enables AI training and real-time LLM inference on models with up to 10 trillion parameters。Nvidia says these technologies will enable breakthroughs in areas such as data processing, engineering simulation, electronic design automation, computer-aided drug design, quantum computing and generative AI。

Super large scale super chip

Nvidia also introduced the GB200 Grace Blackwell super chip。

According to the official, the super chip is connected to two Nvidia B200 Tensor Core GPUs and Grace CPUs through 900GB / s ultra-low power inter-chip interconnection。

GB200 is a key component of Nvidia GB200 NVL72。GB200 NFL72 is a multi-node liquid-cooled rack-level expansion system for highly compute-intensive workloads。It combines 36 GraceBlackwell superchips, including 72 Blackwell GPUs and 36 Grace CPUs connected to each other through the fifth-generation NVLink。

In addition, the GB200 NVL72 has a built-in BlueField ® -3 data processor for cloud network acceleration, combined storage, zero-trust security, and GPU compute elasticity in a hyperscale AI cloud。For LLM inference workloads, compared to the same number of H100s, the GB200 NVL72 can provide up to 30 times the performance improvement and up to 25 times the cost and energy consumption reduction。

It is worth mentioning that the platform is available as a single GPU with 1.4 exaflops of AI performance and 30TB of fast memory are the basic modules that make up the latest generation of DGX SuperPOD。

There are a number of enterprises "desire"Blackwell Architecture

Nvidia revealed that its partners will start supplying products with Blackwell architecture later this year.。

Amazon's AWS, Google Cloud, Microsoft's Azure, and Oracle Cloud Infrastructure will be the first cloud service providers to offer Blackwell-powered instances.。

Sovereign AI clouds such as Indosat Ooredoo Hutchinson, Nebius, Nexgen Cloud, Oracle EU Sovereign Cloud, Oracle US / UK / Australian Government Cloud, Singapore Telecom will also provide cloud services and infrastructure based on Blackwell architecture。

GB200 will also be available through NVIDIA DGX ™ Cloud, an AI platform designed with leading cloud service providers to give enterprise developers specialized infrastructure and software access to build and deploy advanced generative AI models。AWS, Google Cloud, and Oracle Cloud Infrastructure plan to host new instances with Grace Blackwell later this year。

In addition, Cisco, Dell, Lenovo and Ultra Micro are expected to offer a variety of servers built on Blackwell products。Yongqing Electronics, Asus, Foxconn, Jijia, Pegatron and many other companies will also provide Blackwell-based servers.。

In addition, a growing network of software manufacturers, including Ansys, Cadence and Synopsys, global leaders in engineering simulation, will use Blackwell-based processors to accelerate their software for designing and simulating electrical, mechanical and manufacturing systems and parts。Their customers can use generative AI and accelerated computing at lower cost and more energy efficiency to accelerate time to market。

Generative AI microservices

During the GTC conference, Nvidia launched dozens of enterprise-grade generative AI microservices that companies can use to create and deploy custom applications on their own platforms, while retaining full ownership and control of intellectual property.。

Leading application, data and cybersecurity platform providers including Adobe, Cadence, CrowdStrike, Getty Images and SAP have pioneered the use of Nvidia AI Enterprise 5.These new generative AI microservices available in 0。

Nvidia founder and CEO Huang Renxun said: "Mature enterprise platforms are sitting on a gold mine of data that can be transformed into generative AI assistants.。These containerized AI microservices, which we have created with our partner ecosystem, are the building blocks for companies in all industries to become AI companies.。"

NIM Microservices provides pre-built containers based on Nvidia Inference Software, enabling developers to reduce deployment time from weeks to minutes。They provide industry-standard APIs for areas such as language, voice, and drug discovery, enabling developers to quickly build AI applications using proprietary data securely hosted in their own infrastructure.。

NIM microservices will provide fast and high-performance production-grade AI containers for open models of Google, Meta, Microsoft and many other enterprises。

In addition, to accelerate AI adoption, enterprises can use CUDA-X microservices, including Riva for customized voice and translation AI, cuOpt ™ for route optimization, and Earth-2 for high-resolution climate and weather simulation.。

Nvidia says more NVIDIA NeMo ™ microservices for custom model development are coming soon, including NeMo Curator for building concise datasets for training and retrieval, NeMo Customizer for fine-tuning LLM with domain-specific data, NeMo Evaluator for analyzing AI model performance, and NeMo Guardrails for LLM。

General Basic Model of Humanoid Robot - - Project GR00T

At GTC, Nvidia released Project GR00T, a general-purpose basic model for humanoid robots, demonstrating its breakthroughs in robotics and embodied intelligence.。

According to officials, GR00T-powered robots will be able to understand natural language and mimic movements by observing human behavior - quickly learning coordination, flexibility and other skills to navigate, adapt and interact with the real world.。

Nvidia also released a new humanoid robot computer Jetson Thor based on Thor System-on-Chip (SoC)。

The SoC includes a next-generation GPU with transformer engine, which uses Blackwell architecture to provide 800 trillion 8-bit floating-point operations per second AI performance to run multi-modal generation AI models such as GR00T。Dramatically simplifies design and integration with integrated functional safety processor, high-performance CPU cluster, and 100GB Ethernet bandwidth。

Nvidia has revealed that it is developing a comprehensive AI platform for leading humanoid robotics companies such as 1X Technologies, Agility Robotics, Boston Dynamics, Figure AI, and Utrex Technology.。

In addition, Nvidia has also made a major upgrade to the Isaac robot platform.。The Isaac tool used by GR00T is also capable of creating new base models for any robot in any environment。These tools include Isaac Lab for reinforcement learning and OSMO for computational orchestration services。However, the new Isaac platform features won't be available until next quarter.。