Briefly: Nvidia is reportedly considering a transfer to a socketed design for its next-generation GB300 chips. AMD carried out an analogous change, but when Crew Inexperienced throws its appreciable weight behind this configuration, it may set a brand new normal for high-performance AI accelerators.
Nvidia is considering a big change in its design strategy for the upcoming GB300 AI chips. The corporate is exploring the potential for adopting a socketed design for these next-generation merchandise, a departure from its present follow of mounting GPUs straight onto motherboards, in line with reviews from MoneyDJ and the Financial Every day Information that have been noticed by Trendforce.
The GB300, an improve to the lately launched GB200 Blackwell AI {hardware}, is predicted to debut in 2025. If carried out, this design change would mark a primary for Nvidia in its GPU merchandise. The proposed socketed design would contain a motherboard with one CPU and 4 GPU sockets, doubling the variety of GPUs in comparison with the present GB200 mannequin. This is able to permit particular person substitute or improve functionality for every GPU.
There are each advantages and disadvantages to this potential shift. The benefits embrace improved motherboard yields, a simplified provide chain for Nvidia and companions, simpler post-sale help and upkeep, and the potential for future upgrades with out changing the complete board. The challenges embrace potential discount in total platform efficiency and elevated energy and cooling calls for.
The transfer to a socketed design would additionally profit Nvidia’s manufacturing companions, similar to Foxconn. AI server producers won’t must be geared up with SMT manufacturing traces, and there could possibly be a possible enhance for Taiwanese provide chain firms like LOTES.
Whereas this is able to be a primary for Nvidia, rival AMD launched a socketed design in 2023 with its Intuition MI300A AI Accelerator, which makes use of an SH5 socket just like its SP5 socket for Epyc server CPUs.
However with Nvidia now considering this variation, it raises questions on the way forward for GPU design. It is unlikely to have an effect on shopper GeForce playing cards as a consequence of completely different reminiscence configurations. Nonetheless, it may affect different AI chip producers like Intel and should result in new cooling options for high-performance GPUs.
The rumored change has sparked discussions within the tech trade. Analysts recommend it may simplify after-sales service and server board upkeep, whereas some speculate it’d optimize computing board manufacturing yields. Others warning about potential efficiency trade-offs.
Whether or not it encompasses a socketed design or not, the B300 sequence is predicted to change into Nvidia’s mainstream product within the second half of 2025, with demand anticipated to be important as a consequence of its adoption of FP4, which is well-suited for inference eventualities.
In the meantime, Nvidia’s present Blackwell GPUs are experiencing a 12-month backlog fueled by main tech giants like AWS, CoreWeave, Google, Meta, Microsoft, and Oracle. These firms have bought each Blackwell GPU Nvidia and its manufacturing companion TSMC can produce for the following 4 quarters.