Developing Ethernet! NVIDIA is very serious

Recently, at the Computex exhibition in Taiwan, NVIDIA CEO Jensen Huang detailed the innovation of the Spectrum-X platform.

Traditional Ethernet has long been considered unsuitable for use in artificial intelligence due to issues such as slow speed and data loss.

Spectrum-X is the world's first Ethernet architecture designed specifically for AI workloads, capable of improving the performance of generative AI networks by 1.6 times compared to traditional Ethernet. This platform integrates Nvidia Spectrum X Ethernet switches and Nvidia BlueField-3 NICs to address the bottlenecks of traditional Ethernet in AI applications.

Jensen Huang pointed out that traditional Ethernet design focuses on high average throughput, but in deep learning and AI use cases, GPUs mainly communicate with each other, requiring different demands. The Spectrum-X architecture integrates advanced network-level Remote Direct Memory Access (RDMA) and enhanced congestion control functions, significantly improving performance and avoiding hot spot issues. Adaptive routing and noise isolation technologies further optimize Ethernet data transmission efficiency and reliability.

In recent years, NVIDIA's progress in the field of artificial intelligence has been remarkable.

Since acquiring Mellanox for $6.9 billion in 2019, NVIDIA has been advancing its Infiniband technology. Infiniband is a computer network communication standard used for high-performance computing, with extremely high throughput and low latency, used for interconnecting data between computers. In the era of generative AI, it has been widely applied in cloud providers' data centers.

However, not every data center can handle Infiniband, and Ethernet remains the default setting for most data centers. Therefore, there is still a demand for traditional Ethernet that will not change.

To address this, NVIDIA has introduced the Spectrum-X Ethernet platform to help overcome the shortcomings of traditional Ethernet in AI use cases and achieve breakthrough development in artificial intelligence networks.

Market demand has also prompted NVIDIA to invest more resources in Ethernet: according to NVIDIA's financial report for the first quarter of the 2025 fiscal year, the sales of the data center division have grown significantly to $22.56 billion, a 5.3-fold increase year-on-year. However, the supply of Infiniband products has not been able to meet the demand, leading to relatively slow growth in network business revenue. In contrast, Spectrum-X Ethernet products are widely adopted by hyperscale enterprises and cloud builders, becoming the new favorite of data centers.

Chief Financial Officer Collette Kress stated that Spectrum-X sales continue to grow, with customers including large clusters with 100,000 GPUs. It is expected that Spectrum-X will become a multi-billion dollar product line within a year, driving further expansion of NVIDIA in the Ethernet market.

Despite the potential cannibalization effect of Ethernet development on Infiniband sales, Nvidia is actively addressing this challenge.

Jensen Huang emphasized in the conference call that NVIDIA is committed to advancing the development of three computing architectures: NVLink architecture, Infiniband architecture, and Ethernet architecture. The next generation of CPUs, GPUs, network NICs, and switches are about to be launched, all of which will run NVIDIA's CUDA software stack.

Huang said, "We will drive the development of these three areas at a very fast pace. You will see new switches, new NICs, new features, and new software stacks running on these three areas. If you invest in our architecture today, you don't need to do anything, it will enter more and more clouds and data centers, and everything will run smoothly."