NVIDIA Unveils Blackwell B300 GPU

The Blackwell B300 isn't just a single chip; it's the high-end component of a broader platform. The architecture is also featured in the B200 and the GB200 "Superchip," which combines two Blackwell GPUs with a Grace CPU. This variety allows for different system configurations, from single servers to massive, liquid-cooled, rack-scale systems like the GB200 NVL72, which links 72 Blackwell GPUs to function as one giant GPU. A key enabler for these large-scale systems is the fifth generation of NVLink. This interconnect provides 1.8 TB/s of GPU-to-GPU bandwidth, which is double the previous generation and over 14 times faster than PCIe 5.0. This massive bandwidth is critical for training trillion-parameter models, as it allows hundreds of GPUs to communicate with extremely low latency. The move to silicon photonics is a direct response to the physical limitations of copper wiring in data centers. As AI clusters scale, moving data between chips using traditional electrical signals consumes enormous amounts of power and hits bandwidth limits. By converting electrical signals to light (photons), data can travel through fiber optic cables with negligible power loss, drastically improving energy efficiency and enabling faster data transfer. NVIDIA's $4 billion investment is spread across Coherent and Lumentum, two leaders in optical components. This isn't just a financial stake; the deals include multi-year purchase commitments and priority access to manufacturing capacity for advanced lasers and optical networking products. This strategic move helps NVIDIA avoid manufacturing bottlenecks and vertically integrate a critical part of the AI hardware stack, extending from chips to the networking fabric.