Announcement of the next generation of artificial intelligence infrastructureJericho4 Ethernet Switch Router PlatformLater, at the 2025 OCP APAC Summit (2025 Open Compute Project Asia Pacific Summit) in Taipei, Broadcom Senior Vice President and General Manager of the Core Switch Division Ram Velaga further explained the advantages of an open architecture network platform and emphasized that it can surpass NVIDIA's own NVLink technology and subsequentAcquisition of MellanoxThe network architecture built with InfiniBand and Spectrum technologies is more competitive.
Ram Velaga said that in the past, large-scale distributed computing architectures were mainly used for complex computing, energy exploration, or large-scale data search services such as Google Search. However, they are now widely used in large-scale artificial intelligence computing, corresponding to a scale of more than 100 million accelerators and supporting the operation of large language models.
In computing architecture, the current mainstream approach uses GPUs as accelerators, stacking them to improve the performance of a single computing cluster. Network connections are then used to horizontally scale computing, creating even larger computing resources. However, while NVIDIA primarily uses GPU acceleration, many accelerated computing systems in the market utilize CPU stacking, or even ASIC-based computing components (such as Google's TPU), to meet diverse computing deployment requirements.
Compared to NVIDIA, which uses its NVLink technology to connect its GPUs with different CPUs to achieve vertical acceleration of computing performance, and horizontally scales computing scale through InfiniBand and Spectrum technologies, Ram Velaga emphasized that Broadcom prioritizes using open-architecture Ethernet systems to build more flexible and low-latency accelerated computing systems while avoiding the constraints of specific vendors' operating methods.
Ram Velaga pointed out that even though NVIDIA announced at Computex 2025 this year that it would further open up its cooperation with more computing platforms,NVLink Fusion, claiming to enable partners to use their preferred and designed computing platforms in combination with NVIDIA's GPUs to achieve higher computing performance. However, overall, this is still limited by how much resources NVIDIA is willing to open up. Therefore, it emphasizes that a fully open Ethernet network system designed with industry standards can promote a larger accelerated computing ecosystem.
In further explanation, Ram Velaga stated that even with lower latency under NVLink technology, Broadcom's recently proposed Tomahawk Ultra switch processing chip can also achieve very low data transmission latency performance. At the same time, it can connect more accelerators and the corresponding transmission bandwidth performance is also higher. In addition, the use of an open-architecture Ethernet network system will enable the construction of ultra-large-scale accelerated computing systems without being restricted by specific manufacturer technology.
Ram Velaga believes that Ethernet's open architecture will be the foundation for the future development of accelerated computing. It will also make it easier to connect different accelerators, allowing for flexible deployment to meet diverse computing needs, rather than being limited to selecting a few technical resources from specific vendors.
On the other hand, Ram Velaga also emphasized that Broadcom's products provide higher transmission bandwidth during data exchange, thereby further simplifying the switching path of the network system, which means that the overall transmission delay can be greatly reduced and data transmission congestion problems can be reduced. At the same time, by simplifying the transmission path, the costs generated by line distribution and the number of switches can be reduced, making it easier for users to integrate small data centers scattered across the country into large computing clusters through network connections.
At the same time, Ram Velaga also stated that Broadcom will not cooperate with specific companies, emphasizing that it will cooperate with various companies by adhering to the spirit of open source and responding to the accelerated computing components and model computing needs adopted by different companies.
