NVIDIA is reportedly developing the second generation of Blackwell GPUs, and new rumors suggest that the B300 could achieve a 50% performance boost compared to the current B200 generation, with an increase of only 200W in power consumption, according to a report from SemiAnalysis.
The upcoming GPU, possibly named “Blackwell Ultra”, is expected to focus on AI inference workloads and continue using TSMC’s 4nm lithography. Specifically, it will leverage the 4NP (NVIDIA Performance) process, which is optimized to enhance the efficiency and performance of NVIDIA’s chips. This advancement could significantly bolster AI applications by offering greater computational power while maintaining relatively efficient energy usage.
Increased performance x consumption
The rumored NVIDIA B300 chip is expected to consume 1,400W while delivering a 50% performance increase over its predecessor, the B200, representing a relatively modest 200W additional power draw for such a substantial gain. This efficiency is particularly noteworthy in the data center segment, where power consumption is inherently high.
The upcoming AI-focused B300 GPUs are also speculated to feature 288 GB of 12-Hi HBM3E memory, providing an impressive 8 TB/s bandwidth. This setup promises to enhance computational capacity significantly, enabling faster AI training and inference processes while offering up to 3x cost efficiency.
Another intriguing detail in the rumor is that NVIDIA may offer B300 chips separately from the motherboard and pod. This shift could give enterprises greater flexibility in integrating the chips into their custom-built systems, optimizing for specific workloads.
The NVIDIA B300 series is anticipated to debut in 2025, aiming to drive transformative advancements in server workloads and large-scale AI training, potentially setting new standards for the industry.
