Nvidia's upcoming B300-series processors will leverage TSMC's advanced 4NP fabrication process, promising a significant 50% increase in compute performance with a total design power of 1,400W, which is 200W higher than the previous B200 series.

In a strategic pivot, Nvidia will shift its sales focus for the B300 series, opting to sell components like the B300 on an SXM Puck module instead of entire server systems, thereby expanding its role in the supply chain.

The rollout of Nvidia's first-generation Blackwell B200-series processors has faced challenges, including yield issues and unverified overheating reports that have affected server performance.

The B300 processors will feature 12-Hi HBM3E memory stacks, providing 288 GB of memory and 8 TB/s bandwidth, which will facilitate faster AI training and inference while reducing costs.

To enhance server scalability, Nvidia plans to integrate the new 800G ConnectX-8 network interface card into the B300, which will offer double the bandwidth of the previous 400G ConnectX-7 NIC and additional PCIe lanes.

With the enhancements in the B300 series, users can expect larger batch sizes and extended sequence lengths, which will significantly alleviate latency issues in user interactions.

The new design model for the B300 and GB300 will provide greater flexibility for partners among hyperscalers and OEMs, potentially impacting the pricing and performance of Blackwell-based systems.

Nvidia's second-generation Blackwell B300-series processors are anticipated to launch approximately six months following the introduction of the B200 series.