NVIDIA CEO Jensen Huang announced the company's next-generation GPU architecture, named Blackwell, at the NVIDIA GPU Technology Conference (GTC) held in San Jose, California. This architecture is expected to make significant advancements in the field of artificial intelligence (AI), highlighting the chip manufacturer's continued dominance in the AI hardware market.
The Blackwell GPU is named after renowned mathematician and statistician David Harold Blackwell and is expected to replace the highly successful Hopper architecture introduced two years ago. The Blackwell chip features an impressive 20.8 billion transistors and is manufactured using a customized 4NP TSMC process, resulting in unparalleled performance and efficiency.
A key innovation of the Blackwell architecture is the second-generation Transformer engine, which supports twice the computation and model size compared to its predecessor. This advancement, combined with NVIDIA's TensorRT-LLM and NeMo Megatron frameworks, will enable the development and deployment of trillion-parameter-scale AI models at lower costs and energy consumption.
Leading technology giants including Amazon Web Services, Dell Technologies, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and xAI are expected to adopt the Blackwell platform, further solidifying NVIDIA's position as the preferred AI hardware supplier. These companies have shown enthusiasm for the breakthrough performance capabilities of the Blackwell GPU, which will accelerate discoveries and drive innovation across various industries.
The Blackwell architecture also introduces several other groundbreaking technologies, such as the fifth-generation NVLink, which achieves an astonishing bidirectional throughput of 1.8TB/s per GPU, ensuring seamless communication among hundreds of GPUs to support the most complex large language models (LLMs). Additionally, the RAS engine and advanced confidential computing capabilities enhance system uptime, recovery, and data protection without sacrificing performance.
NVIDIA is dedicated to accelerating data processing and has equipped the Blackwell architecture with a dedicated decompression engine that supports the latest formats and accelerates database queries. This development is particularly important as data processing is expected to increasingly rely on GPU acceleration in the coming years.
The GB200 Grace Blackwell Superchip is a multi-node, liquid-cooled, rack-scale system that integrates 36 Grace Blackwell Superchips. Compared to an equal number of NVIDIA H100 Tensor Core GPUs, it achieves up to 30 times higher performance in large language model (LLM) inference workloads. As a single GPU, this platform delivers 1.4 exaflops of AI performance and 30TB of fast memory, serving as a foundational building block for the latest DGX SuperPOD.
Products based on Blackwell will be offered later this year by a range of partners, including cloud service providers, server manufacturers, and software makers. This global partner network will enable organizations across industries to leverage the powerful capabilities of Blackwell for applications such as generative AI, engineering simulation, electronic design automation, and computer-aided drug design.
As Jensen Huang stated in his keynote speech, "Generative AI is the defining technology of our time. Blackwell is the engine driving this new industrial revolution." With the launch of the Blackwell GPU, NVIDIA is poised to maintain its leadership position in the AI hardware market and drive unprecedented advancements in artificial intelligence and accelerated computing.
