무슨 발표인가
- NVIDIA Grace Blackwell 200 Azure 통합
- 16TB/s 메모리 대역폭
- Hopper 세대 대비 30배 추론 성능
원문 (영어)
At NVIDIA GTC , Microsoft and NVIDIA are announcing new offerings across a breadth of solution areas from leading AI infrastructure to new platform integrations, and industry breakthroughs. Today’s news expands our long-standing collaboration, which has paved the way for revolutionary AI innovations that customers are now bringing to fruition.
Microsoft and NVIDIA collaborate on Grace Blackwell 200 Superchip for next-generation AI models Microsoft and NVIDIA are bringing the power of the NVIDIA Grace Blackwell 200 (GB200) Superchip to Microsoft Azure. The GB200 is a new processor designed specifically for large-scale generative AI workloads, data processing, and high performance workloads, featuring up to a massive 16 TB/s of memory bandwidth and up to an estimated 30 times the inference on trillion parameter models relative to the previous Hopper generation of servers.
Microsoft has worked closely with NVIDIA to ensure their GPUs, including the GB200, can handle the latest large language models (LLMs) trained on Azure AI infrastructure. These models require enormous amounts of data and compute to train and run, and the GB200 will enable Microsoft to help customers scale these resources to new levels of performance and accuracy.
Microsoft will also deploy an end-to-end AI compute fabric with the recently announced NVIDIA Quantum-X800 InfiniBand networking platform. By taking advantage of its in-network computing capabilities with SHARPv4, and its added support for FP8 for leading-edge AI techniques, NVIDIA Quantum-X800 extends the GB200’s parallel computing tasks into massive GPU scale.
원문: Microsoft Source — "Microsoft and NVIDIA continue to deliver on the promise of AI | Microsoft Azure Blog" (2024-03-18) 공식 원문: https://news.microsoft.com/2024/03/18/microsoft-and-nvidia-announce-major-integrations-to-accelerate-generative-ai-for-enterprises-everywhere/