NVIDIA (NVDA US) GB200 vs. GB300? – The Naming Game
NVIDIA has been making waves recently with a flurry of changes. It is said that the company has cancelled the NVL36 and B200A, while introducing new models like the B300, GB300, and GB300A. The flood of product names has left some people unfamiliar with NVIDIA's product portfolio feeling confused. This article aims to help general investors get a clearer understanding of the recent product changes.
First, we know that there are two models that will undergo production start in December this year: the B200 and GB200. The B200 is a standard HGX server composed of 8 Blackwell GPUs. Each Blackwell consists of 2 GPU dies and 8 eight-layer stacked HBM3e dies, then packaged by using CoWoS-L process (for more details on the various types of CoWoS advanced packaging, please refer to my previous article: The Bonder War – An update on ASMPT (522 HK)，BESI (BESI NA), Shibaura (6590 JP), Tazmo (6266 JP) and other advanced packaging equipment players). The GB200 adds Grace CPU in addition to Blackwell GPU. Each of GB200 compute tray contains 2 Grace CPUs and 4 Blackwell GPUs, then connected through NVLink into a full rack system. The system comes in two configurations: NVL36 (9 compute trays with a total of 36 GPUs) and NVL72 (18 compute trays with 72 GPUs). NVIDIA currently recommends customers to adopt the NVL72 configuration, while they will still support to ship the NVL36 systems for customers who have already initiated their NVL36 cases (for a detailed introduction of the GB200 system, please refer to my previous article: Macom (MTSI US) – A Hidden Nvidia GB200 Play).
Next, let's talk about the two models that NVIDIA has recently renamed: the B300A and B300.
The B300A was originally called the B200A Ultra. It differs from the B200 which was mentioned above in the following 3 ways:
The "A" suffix indicates that the Blackwell GPU is made up of 1 GPU die and 4 HBM3e dies (vs. 2 GPU dies and 8 HBM3e dies in the B200).
Since the B300A only has one GPU die, it uses the conventional CoWoS-S packaging (vs. B200 adopts CoWoS-L packaging which has larger interposer area to support two GPU dies).
The "Ultra" in the original name meant that it uses 12-layer stacked HBM3e dies (vs. 8-layer stacked HBM3e dies in the B200), each HBM cube then has 36GB of memory capacity instead of 24GB.
The CoWoS-S-based B200A chip was originally expected to be ready by the end of this year, but after being renamed, they are now scheduled for production start in June next year. Moreover, they will no longer be sold as standard 8-GPU HGX servers but in the format of GB300A NVL36 rack system. As a result, TSMC has recently increased its CoWoS-S capacity plan for 2H25 from 24kwpm to 28kwpm, in preparation for the GB300A production in Q325. The GB300A system also has a unique design where each compute tray contains only 1 Grace CPU paired with 4 Blackwell GPUs (vs. the original design of 2 Grace CPUs paired with 4 Blackwell GPUs), which results in meaningful power savings (0.7kw vs.1.2kw of GB200).
The B300 was originally called the B200 Ultra, which, as the name suggests, has 12-layer stacked HBM3e dies (vs. 8-layer stacked HBM3e dies in the B200). It will be available in both standard 8-GPU HGX server format and NVL72 rack system format (i.e., GB300). Both the B300 and GB300 are scheduled for production start in June next year.
Finally, I’ve compiled the following table to help everyone keep track of NVIDIA's upcoming product roadmap:
