A significant change could be on the horizon for NVIDIA’s AI server landscape. According to recent reports, the company is rumored to be moving towards a socketed design for its next-generation GB300 AI servers, powered by the upcoming Blackwell Ultra AI chips slated for release in 2025. This shift marks a departure from NVIDIA’s traditional approach and could have major implications for the industry.
TrendForce, a leading industry research firm, suggests that the B300 series will become NVIDIA’s mainstream offering in the second half of 2025. The key driver behind this move appears to be the adoption of FP4 technology, which is highly efficient in inference scenarios. The socketed design is also expected to boost the yield rates of the B300 AI GPUs, although TrendForce acknowledges that it might lead to a slight performance reduction.
The Economic Daily News further highlights the advantages of this design change, stating that it will simplify after-sales service and server board maintenance. The socketed design also optimizes the yield of computing board manufacturing. This flexibility allows AI server manufacturers to potentially operate without needing an SMT production line, further streamlining the production process.
If these rumors prove true, it would be a first for NVIDIA, marking a departure from its past GPU designs. In contrast, competitor AMD already embraced a socket design in 2023 with its Instinct MI300A AI Accelerator. The move by NVIDIA suggests a potential shift in the industry towards a more modular approach, allowing for greater flexibility and potentially driving innovation in AI server development.
The introduction of socketed designs in NVIDIA’s AI server offerings could have a significant impact on the industry. While the potential performance trade-off needs to be considered, the advantages in terms of production, maintenance, and cost optimization make it a compelling strategy for NVIDIA. As we approach the launch of the B300 series in 2025, it will be interesting to see how this shift shapes the future of AI server technology.
