Nvidia Corporation, a renowned technology company, announced the full production of its groundbreaking GH200 Grace Hopper Superchip, touted as the company’s most powerful artificial intelligence chip to date.
Revealed by Nvidia CEO Jensen Huang at the COMPUTEX 2023 event in Taiwan, the GH200 Superchip is designed to fuel systems capable of handling intricate AI workloads, including training the next generation of generative AI models.
Combining Nvidia’s Arm-based Grace central processing unit and Hopper graphics processing unit architectures, the GH200 integrates these powerful components into a single chip using Nvidia’s NVLink-C2C interconnect technology.
With an impressive total bandwidth of 900 gigabytes per second, the Superchip outperforms today’s leading accelerated computing systems, offering seven times more bandwidth than standard PCIe Gen5 lanes. Nvidia also highlights its power efficiency, consuming five times less power, which enables it to meet the demands of resource-intensive AI and high-powered computing applications more effectively.
The GH200 Superchip holds particular promise for generative AI workloads, exemplified by OpenAI LP’s ChatGPT. This groundbreaking technology has captivated the industry with its nearly human-like ability to generate content from prompts.
Ian Buck, Nvidia’s vice president of Accelerated Computing, commented on the significance of generative AI, stating, “Generative AI is rapidly transforming businesses, unlocking new opportunities and accelerating discovery in healthcare, finance, business services, and many more industries.” He further added that the GH200 Superchips will provide the accelerated infrastructure needed to develop and deploy generative AI applications that leverage proprietary data.
The first system to incorporate the GH200 Superchip will be Nvidia’s own next-generation AI supercomputer, the Nvidia DGX GH200. This state-of-the-art system utilizes the NVLink Switch System to combine 256 GH200 superchips, resulting in a single GPU with an impressive performance of 1 exaflop, or 1 quintillion floating-point operations per second, and shared memory of 144 terabytes.
Compared to Nvidia’s previous DGX A100 supercomputer, the DGX GH200 offers nearly 500 times more memory and significantly enhanced processing power. The DGX GH200 AI supercomputers will be equipped with full-stack software, including Nvidia’s Base Command software for AI workflow management and Nvidia AI Enterprise software, which offers over 100 AI frameworks, pretrained models, and development tools for efficient generative AI, computer vision, and speech AI model production.
Google Cloud, Meta Platforms Inc., and Microsoft Corp. are among the first customers set to benefit from the DGX GH200 AI supercomputers. Nvidia also plans to provide the DGX GH200 design as a blueprint for cloud service providers to customize according to their infrastructure requirements.
Microsoft’s Girish Bablani, corporate vice president of Azure Infrastructure, emphasized the potential of the DGX GH200, stating, “The potential for DGX GH200 to work with terabyte-sized datasets would allow developers to conduct advanced research at a larger scale and accelerated speeds.”
With the GH200 Superchip, Nvidia continues to push the boundaries of AI and computing, offering enhanced capabilities that are set to transform various industries and drive innovation forward.