The NVIDIA B300 is the next-generation Blackwell architecture GPU announced for release in 2025, representing the successor to the B200 with further expanded memory capacity. It is expected to feature 192GB of HBM3e memory, providing the largest memory capacity of any single NVIDIA datacenter GPU to date.
The B300 continues the Blackwell architectural innovations introduced with B200, including the dual-die design with two GPU chiplets connected by a high-bandwidth interconnect on a single package. This chiplet approach allows NVIDIA to scale compute resources beyond the reticle limits of a single die while maintaining coherent memory access.
Expected specifications include 5th-generation NVLink with increased bandwidth over B200, 5th-generation Tensor Cores with enhanced FP4 and FP8 support, and next-generation Transformer Engine optimizations. The SXM form factor will support NVL72 configurations enabling 72-GPU fully-connected domains.
The 192GB memory capacity targets training and inference of trillion-parameter models and beyond. Combined with expected memory bandwidth improvements, the B300 addresses the growing model size requirements of frontier AI systems. TDP is expected in the 1000W+ range, requiring advanced liquid cooling solutions.