For the new (G)B300 super accelerators, NVIDIA has significantly changed the supply chain, making it more friendly to hyperscalers, that is, the main customers of new products, reports SemiAnalysis. In the case of the GB200, the company supplied ready-made, fully integrated Bianca boards, including Blackwell accelerators, CPU Grace, 512 GB of soldered LPDDR5X, VRM, etc. GB300 will be supplied in the form of modules: SXM Puck B300, CPU Grace in a BGA package, HMC from Axiado (instead of Aspeed). And LPCAMM modules, mainly from Micron, will be used as system RAM.
The transition to SXM Puck will enable more OEM and ODM suppliers, as well as hyperscalers themselves, to create new systems. If previously only Wistron and Foxconn could produce Bianca boards, now others can join the accelerator assembly process. Wistron stands to lose the most from this decision as it will lose market share from Bianca manufacturers. For Foxconn, which thanks to NVIDIA is about to become the world’s largest server supplier, the loss is compensated by the exclusive production of SXM Puck.
Image source: NVIDIA
Another important change concerns VRM. While the SXM Puck has some VRM components, most of the remaining components will be purchased by hyperscalers and vendors directly from VRM suppliers. Rackmount NVSwitches and copper backplane will continue to be supplied by NVIDIA itself. For the GB300, the company offers the 800G InfiniBand/Ethernet Quantum-X800/Spectrum-X800 platform with ConnectX-8 adapters, which were not included in the GB200 due to inconsistencies in the timing of product launches. In addition, ConnectX-8 has 48 PCIe 6.0 lanes at once, which allows you to create unique architectures such as the air-cooled MGX B300A.
It is reported that all key hyperscalers have already decided to move to GB300. This is partly due to the higher performance and efficiency of the GB300, but also due to the fact that they can now customize the platform, cooling system, etc. For example, Amazon could finally use its own water-cooled motherboard and return to the NVL72 architecture, improving TCO. Previously, the company was the only major player to choose the less efficient NVL36 option due to the use of its own 200G adapters and air-cooled PCIe switches.
However, there is a drawback – hyperscalers will have to spend more time and resources on designing and testing the product. This is perhaps the most complex platform that hyperscalers have ever had to design (with the exception of Google TPU platforms), noted the SemiAnalysis resource.