Nvidia
On Wednesday, Nvidia introduced a collaboration with Microsoft to construct a “huge” cloud pc targeted on AI. It’ll reportedly use tens of 1000’s of high-end Nvidia GPUs for purposes like deep studying and enormous language fashions. The businesses intention to make it probably the most highly effective AI supercomputers on this planet.
In flip, the brand new supercomputer will characteristic 1000’s of items of what’s arguably probably the most highly effective GPU on this planet, the Hopper H100, which Nvidia launched in October. Nvidia will even present its second strongest GPU, the A100, and make the most of its Quantum-2 InfiniBand networking platform, which may switch knowledge at 400 gigabits per second between servers, linking them collectively into a robust cluster.
In the meantime, Microsoft will contribute its Azure cloud infrastructure and ND- and NC-series digital machines. Nvidia’s AI Enterprise platform will tie the entire thing collectively. The businesses will even collaborate on DeepSpeed, Microsoft’s deep studying optimization software program.
In an announcement, Nvidia talked about the purposes the joint supercomputer may serve:
“As a part of the collaboration, Nvidia will make the most of Azure’s scalable digital machine situations to analysis and additional speed up advances in generative AI, a quickly rising space of AI during which foundational fashions like Megatron Turing NLG 530B are the idea for unsupervised, self-learning algorithms to create new textual content, code, digital pictures, video or audio.”
This previous yr has seen a fast rise in generative AI fashions corresponding to Secure Diffusion and DALL-E that may synthesize novel pictures on demand. Related fashions have appeared that may create video, synthesize voices, and carry out transcription, amongst different makes use of. As computational demand will increase for generative AI, Nvidia and Microsoft intend to be there to fulfill it.
As soon as Nvidia and Microsoft’s cloud pc comes on-line, clients can deploy 1000’s of GPUs in a single cluster to “practice even probably the most huge massive language fashions, construct probably the most complicated recommender methods at scale, and allow generative AI at scale,” in accordance with Nvidia.
The businesses didn’t present particulars on when the brand new supercomputer will likely be prepared however talked about that the announcement marks the start of a “multi-year collaboration.” It is seemingly the cloud pc will scale up in capability over time.