Nvidia is set to rebrand its Blackwell Ultra product line — which will be a refresh kind of family with increased performance — to the B300 series to better differentiate it from the upcoming B100 and B200 products, reports TrendForce. But the upgraded memory configuration and performance will come at a price, according to the report.
What was supposedly going to be the Nvidia B200 Ultra will be rebranded as the B300, while the GB200 Ultra becomes the GB300. Additional updates include renaming the B200A Ultra to B300A and the GB200A Ultra to GB300A.
Keep in mind that this information is unofficial — Nvidia's previously published roadmap did list "Blackwell Ultra" as a product family, but it never went into specific product names. Whether Blackwell Ultra was going to be called Blackwell B200 Ultra or something else was never announced.
Nvidia's B100, B200, and GB200 come with 192GB of HBM3E memory using 8-Hi HBM3E memory stacks. By contrast, B300 and GB300 will come equipped with 288GB of HBM3E using eight 12-Hi stacks. Since the B300-series products will come with more memory (and possibly with higher-clocked GPUs and/or more processing cores), they are meant to be more expensive than B100 and B200 solutions. By contrast, the more cost-effective B300A and GB300A will feature 144 GB of memory using four 12-Hi HBM3E stacks, according to analysts from TrendForce. Essentially, this means that most of Nvidia's GPUs will switch to 12-Hi HBM3E, which will require DRAM makers to adjust their HBM3E supply chain.
Shipments of Nvidia's B100/B200 and GB200 series are expected to begin between Q4 2024 and Q1 2025. That represents a delay due to some hardware issues that Nvidia has now addressed, and TrendForce believes that Nvidia will roll out B300-series already in Q2 or Q3 2025. The B300A, designed specifically for server OEMs, will enter mass production in mid-2025. Interestingly, TrendForce claims that Nvidia had originally planned to launch a B200A for OEMs but decided to go straight with the B300A, possibly due to lower-than-expected demand for lower-tier GPUs.
While OEM business is important for Nvidia, the company has said it will prioritize its own NVL rack solutions with 36 or 72 Blackwell GPUs designed for cloud service providers including Amazon Web Services, Meta, and Microsoft. Nvidia's NVL36 and NVL72 machines will feature more expensive GPUs and will also cost more than traditional GPU servers, to some degree because the new Blackwell GPUs consume a lot of power and mandate liquid cooling. Transitioning to such machines will come with higher associated costs as companies will have to move to a more advanced data center infrastructure.
