NVIDIA is preparing to unveil its highly anticipated B300 AI GPU, with plans to launch in Q3 2025 followed by mass production in Q4 2025, according to industry reports.
The B300, successor to the current B200 series, will be officially revealed at NVIDIA's GPU Technology Conference (GTC) in March 2025. The new AI processor is expected to deliver 1.5 times better FP4 performance compared to its predecessor while consuming up to 1400W of power.
A major upgrade comes in memory capacity, with the B300 featuring 288GB of HBM3E memory using a 12-Hi stack configuration - a substantial increase from the B200's 192GB HBM3E with 8-Hi stacks. The platform will also see networking improvements, with optical modules upgraded from 800G to 1.6T speeds.
The B300 will be accompanied by NVIDIA's first CPO technology-based InfiniBand switch, scheduled for Q3 2025. This switch will support 36 3.2T CPO optical modules and four 28.8T application-specific integrated circuits, enabling total signal transmission speeds of 115.2T.
Industry analysts view the B300 as NVIDIA's next major product aimed at maintaining its dominant position in the AI accelerator market. The comprehensive improvements in performance, memory capacity, and networking capabilities position the B300 as a powerful solution for next-generation AI workloads.
The announcement comes as NVIDIA's manufacturing partners prepare for increased production demands. TSMC plans to expand its advanced packaging capacity to 80,000 pieces monthly by Q4 2025 to support the rollout of NVIDIA's future AI processors.