Blackwell
NVIDIA GPU architecture unveiled in 2024, succeeding Hopper and Ada Lovelace. It powers the B200 and B300, the RTX 50 and the RTX PRO 6000.
Blackwell is the name of the GPU architecture NVIDIA unveiled in March 2024, named in honour of the mathematician David Blackwell. It succeeds Hopper for data centers and Ada Lovelace for consumer and professional cards. It is found in the B200 and B300 accelerators, in GeForce RTX 50 series cards and in the RTX PRO 6000 Blackwell.
For data centers, Blackwell combines two large dies in a single package, connected by a very fast link, so they work as one GPU. It introduces the fifth generation of Tensor Cores, with support for FP4 precision, which roughly doubles compute throughput compared with FP8 for inference of large models. NVLink also moves to its fifth generation, twice as fast as on Hopper. The RTX cards of this generation adopt GDDR7 memory.
On a product page, the word Blackwell is not enough: an RTX 5060 Ti and a B300 share the architecture but are in completely different categories. Look at the amount and type of memory (GDDR7 or HBM3e), the bandwidth, the power draw and the form factor (PCIe card or SXM module). To run a model locally, one or two RTX 50 cards are often enough; to serve many users or train models, you need to aim for the B200 or B300.
We offer AI workstations based on RTX 50 and RTX PRO 6000 Blackwell cards, as well as B300 servers on quotation.