Nvidia exploring shift to socketed design for next-gen GB300 AI chips

It wouldn't be first - AMD introduced a socketed design in 2023

Serving tech enthusiasts for over 25 years.
TechSpot means tech analysis and advice you can trust.

TL;DR: Nvidia is reportedly contemplating a move to a socketed design for its next-generation GB300 chips. AMD implemented a similar change for data center processors, but if Team Green throws its considerable weight behind this configuration, it could set a new standard for high-performance AI accelerators.

Nvidia is contemplating a significant change in its design approach for the upcoming GB300 AI chips. The company is exploring the possibility of adopting a socketed design for these next-generation products, a departure from its current practice of mounting GPUs directly onto motherboards, according to reports from MoneyDJ and the Economic Daily News that were spotted by Trendforce.

The GB300, an upgrade to the recently launched GB200 Blackwell AI hardware, is expected to debut in 2025. If implemented, this design change would mark a first for Nvidia in its GPU products. The proposed socketed design would involve a motherboard with one CPU and four GPU sockets, doubling the number of GPUs compared to the current GB200 model. This would allow individual replacement or upgrade capability for each GPU.

There are both benefits and drawbacks to this potential shift. The advantages include improved motherboard yields, a simplified supply chain for Nvidia and partners, easier post-sale support and maintenance, and the potential for future upgrades without replacing the entire board. The challenges include possible reduction in overall platform performance and increased power and cooling demands.

The move to a socketed design would also benefit Nvidia's manufacturing partners, such as Foxconn. AI server manufacturers might not need to be equipped with SMT production lines, and there could be a potential boost for Taiwanese supply chain companies like LOTES.

// Related Stories

• Elon Musk's xAI built a 100,000-GPU supercluster in just 19 days - normally takes years
• AMD's Navi 44 "RDNA 4" GPU package size could be much smaller than Navi 33

While this would be a first for Nvidia, rival AMD introduced a socketed design in 2023 with its Instinct MI300A AI Accelerator, which uses an SH5 socket similar to its SP5 socket for Epyc server CPUs.

But with Nvidia now contemplating this change, it raises questions about the future of GPU design. It's unlikely to affect consumer GeForce cards due to different memory configurations. Still, it could influence other AI chip manufacturers like Intel and may lead to new cooling solutions for high-performance GPUs.

The rumored change has sparked discussions in the tech industry. Analysts suggest it could simplify after-sales service and server board maintenance, while some speculate it might optimize computing board manufacturing yields. Others caution about potential performance trade-offs.

Whether it features a socketed design or not, the B300 series is expected to become Nvidia's mainstream product in the second half of 2025, with demand expected to be significant due to its adoption of FP4, which is well-suited for inference scenarios.

Meanwhile, Nvidia's current Blackwell GPUs are experiencing a 12-month backlog fueled by major tech giants like AWS, CoreWeave, Google, Meta, Microsoft, and Oracle. These companies have purchased every Blackwell GPU Nvidia and its manufacturing partner TSMC can produce for the next four quarters.