Nvidia encountered significant challenges during the launch of its pioneering Blackwell B200-series processors, including production issues that affected yields and reports of servers overheating. Despite these setbacks, as per SemiAnalysis, Nvidia is on the brink of launching its second-generation Blackwell B300-series processors. These new processors promise a 50% boost in performance and increased memory capacity, all while only requiring an additional 200W of thermal design power (TDP).
The B300-series from Nvidia showcases a revamped architecture that will continue to utilize TSMC’s 4NP fabrication process, a 4nm node specially optimized for Nvidia’s requirements. According to the report, the B300-series will deliver 50% more computing power compared to the B200-series. This performance leap will come with a substantial TDP of 1,400W, which is just 200W more than its predecessor, the GB200. The B300 is expected to hit the market about six months after the B200, according to SemiAnalysis.
Another notable upgrade in the B300-series is the integration of 12-Hi HBM3E memory stacks, which will provide a massive 288 GB of memory and 8 TB/s of bandwidth. This significant increase in memory capacity and computational power will enhance the efficiency of training and inference processes, reducing the cost of inference by up to threefold. The B300’s ability to manage larger batch sizes and support extended sequence lengths will also help in minimizing latency during user interactions.
Beyond the enhancements in computational performance and memory, the second-generation Blackwell processors might also include Nvidia’s 800G ConnectX-8 NIC. This network interface card doubles the bandwidth of the existing 400G ConnectX-7 and offers 48 PCIe lanes, up from 32 in the previous generation. This improvement will greatly enhance the scale-out bandwidth capabilities of new servers, benefiting large-scale cluster deployments.
In a strategic shift, Nvidia plans to overhaul its supply chain for the B300 and GB300. Unlike its approach with the B200 and GB200, where it attempted to market the entire reference motherboard or server pod, Nvidia will now only sell the B300 on an SXM Puck module, accompanied by the Grace CPU and a host management controller (HMC) from Axiado. This change will open up the Blackwell supply chain to more participants, potentially making Blackwell-based systems more widely available.
With the introduction of the B300 and GB300, Nvidia is set to offer its partners, including hyperscalers and OEMs, greater flexibility in designing Blackwell-based systems. This could influence both the pricing and performance of these machines.
