Chips, compute and data centres
Nvidia Blackwell
In short
Nvidia Blackwell is a GPU architecture and product family that Nvidia announced in March 2024 for accelerated computing, including AI training and inference. It includes the B200 GPU and racks that pair Blackwell GPUs with Grace CPUs.
- Also called
- Blackwell, NVIDIA Blackwell
- Topic
- Chips, compute and data centres
- Level
- Some background helps
- First described
- Nvidia announced the Blackwell platform at GTC on 18 March 20241.
Nvidia announced the Blackwell platform at GTC on 18 March 2024. The announcement describes a GPU architecture for accelerated computing, including data processing, engineering simulation and generative AI, together with NVLink for connecting processors1.
The announcement lists a second-generation Transformer Engine, fifth-generation NVLink with 1.8 terabytes per second of bidirectional throughput per GPU, and B200 Tensor Core GPUs. A GB200 Grace Blackwell Superchip connects two B200 GPUs to one Grace CPU. Nvidia claims up to 25 times less cost and energy than the previous platform. The same release marks statements about performance and benefits as forward-looking1.
Blackwell is not one board. The announcement also describes the HGX B200, which links eight B200 GPUs, and the GB200 NVL72, a liquid-cooled rack of 36 Grace Blackwell Superchips: 72 Blackwell GPUs and 36 Grace CPUs on fifth-generation NVLink1. Nvidia's DGX GB rack guide says each NVL72 compute tray has two Grace CPUs and four Blackwell GPUs. Liquid runs through manifolds and cold plates on those CPUs and GPUs. Fans cool networking and storage. The guide puts rack power at about 120 kilowatts2.
In May 2026 Nvidia announced that its Vera Rubin platform was ramping into full production, and described Grace Blackwell as the previous generation. It said Vera Rubin production shipments were set to begin in autumn 2026. The DGX GB guide, as fetched in October 2026, still documents both GB200 and GB300 racks and calls their GPUs Blackwell GPUs23.
Example
Nvidia's March 2024 announcement named Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure among the first cloud providers that would offer Blackwell-powered instances1.
Why it matters
Blackwell is the GPU platform Nvidia launched in 2024 for large AI systems, and in May 2026 Nvidia presented Vera Rubin as the following generation. The GB200 rack description also shows the NVLink fabric, the liquid cooling and the roughly 120 kilowatt rack power these designs use123.
Questions people ask
Is Blackwell a GPU or an AI model?
Nvidia's March 2024 announcement describes a GPU architecture and hardware products, including the B200 GPU and the GB200 NVL72 rack. The announcement does not describe an AI model1.
What is a GB200 NVL72?
Nvidia's announcement describes a liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs linked by fifth-generation NVLink. The DGX GB guide says liquid cools those CPUs and GPUs through cold plates, and fans cool networking and storage12.
Is Blackwell still Nvidia's latest GPU platform?
Nvidia's May 2026 announcement calls the Grace Blackwell platform the previous generation next to Vera Rubin, which it said was ramping into full production, with shipments set to begin in autumn 2026. Nvidia's DGX GB guide still documents GB200 and GB300 racks of Blackwell GPUs23.