NVIDIA announced the Blackwell platform, including the GB200 NVL72 rack design. Its NVLink fabric connects 72 Blackwell GPUs so they can exchange intermediate results when a model spreads its layers or experts across devices.
Company / HubLLMImageVideoCross-Modal
NVIDIA
Plain English
NVIDIA designs GPUs used to train and run foundation models. It also builds the links that let many GPUs work on the same model.
Technical Overview
NVIDIA Tensor Cores accelerate the matrix operations used in neural networks. NVLink carries data between GPUs during distributed training and inference. Some GPUs can also be partitioned into isolated instances for separate workloads.
Chronological Entries (2)
#2Milestone
AI-Assisted
#1Release
AI-Assisted
NVIDIA announced that its Ampere A100 GPU was shipping. Its Multi-Instance GPU feature lets an operator divide one physical GPU into as many as seven isolated instances, so smaller inference jobs can share the device with dedicated compute and memory.
Direct Relations (1)
- EnablesDistributed Training