What Is NVIDIA Blackwell?

NVIDIA Blackwell is a GPU architecture: the underlying design of the processing cores, memory system and interconnects that NVIDIA uses across a generation of products. NVIDIA announced it at its GTC conference in March 2024 and named it after David Blackwell, the American mathematician and statistician.

Blackwell replaced two earlier architectures at once. In the data centre it follows Hopper, the architecture of the H100 and H200. In graphics, workstation and embedded products it follows Ada Lovelace, used in the RTX 6000 Ada and GeForce RTX 40 series. That is why the name covers products as different as a liquid-cooled rack of 72 GPUs and a palm-sized robotics module.

What ties them together is the same generation of CUDA cores, Tensor Cores and software support, so a model developed on one Blackwell product can run, with suitable sizing, on another.

The headline changes are aimed at AI, and in particular at running large language and vision models more efficiently. Not every feature appears on every product, because data centre and edge parts are built for very different power budgets.

The same architecture is built into products that differ by a factor of more than ten in power. The table summarises the main lines and where each one is used.

For industrial systems, the relevant parts are the RTX PRO cards and modules and Jetson Thor. Our RTX PRO Blackwell computers range from compact edge systems with low-profile cards to servers with the RTX PRO 6000. Machine vision PCs such as the VCO-6000-RPL series list RTX PRO Blackwell cards up to the RTX PRO 6000 Blackwell Max-Q on their GPU support list, and MXM GPU modules bring the architecture to sealed and conduction-cooled systems with no room for a PCIe card.

Against Hopper in the data centre, Blackwell’s main gains are FP4 support, more memory per GPU and a larger NVLink domain, which together cut the cost of serving large models. Against Ada Lovelace in workstation cards, the change is easiest to see at the top of the range: the RTX PRO 6000 Blackwell has 24,064 CUDA cores and 96 GB of memory, against 18,176 cores and 48 GB on the RTX 6000 Ada.

Software carries over. Blackwell runs current CUDA, TensorRT and the usual AI frameworks, although older code gains most after it is rebuilt for the new architecture and quantised to the lower precisions it supports. Our FAQ on INT8 quantisation explains how lower-precision models trade a little accuracy for speed and memory.

At the edge, the important Blackwell product is Jetson Thor. Its GPU has 2,560 CUDA cores and 96 fifth-generation Tensor Cores and is rated at up to 2,070 FP4 TFLOPS on the top module, several times the AI throughput of Jetson AGX Orin. That makes on-device language models, vision-language models and multi-camera robotics practical within a power budget a vehicle or robot can supply. Our Jetson Thor range includes rugged sealed systems such as the WCO-6000-THR-A series, with 10GbE and GMSL2 camera inputs for sensor-heavy machines.

For larger models or more streams, an x86 system with an RTX PRO Blackwell card gives more memory and throughput, at higher power. The usual choice comes down to the size of the model, the number of video streams and the power and cooling available on site.

We fit and validate RTX PRO Blackwell cards and Jetson Thor modules in our edge AI computers as standard, with the drivers, CUDA version and AI software your models are built against.