NVIDIA has confirmed the operational status of its new Blackwell AI GPUs by running them in real-time within its data centers. This move effectively dispels previous rumors suggesting significant design challenges and issues with the water-cooling systems of Blackwell AI servers.
NVIDIA’s announcement confirms Blackwell’s progress towards production, with limited shipments slated for the fourth quarter of 2024. The company plans to divulge more information about the Blackwell architecture at the upcoming Hot Chips conference next week.
The company also released captivating visuals of various trays within the Blackwell family, offering a glimpse into the complex engineering and design behind these powerful units. The sheer scale and intricate details are remarkable, showcasing the artistry behind the technology.
Delving into the core of the Blackwell B200 AI GPU, we encounter an impressive 208 billion transistors (or 104 billion transistors when two reticle-limited GPUs are combined). This feat is achieved using TSMC’s 4NP process node. The GPU boasts 192GB of HBM3E memory with an astounding 8TB/sec memory bandwidth. Further enhancing its capabilities is a 1.8TB/sec bi-directional NVLink bandwidth, complemented by a high-speed NVLink-C2C Link to the Grace CPU.
