The efficiency and reliability of the global artificial intelligence (AI) and high-performance computing (HPC) military competition have become the decisive factors. Broadcom announced on the 5th that Ethernet has established a leading position in large-scale network solutions, surpassing Infiniband, which was once favored.
Ram Velaga, Vice President of Broadcom and General Manager of the Core Exchanger Department, said Infiniband is known for his agility, but has trust issues. In contrast, Ethernet has become the core of global network information transmission with its highest efficiency and reliability, as well as rapid failure recovery capabilities. Meta and Amhor two major technological giants have further demonstrated the superiority of Ethernet.
Ram Velaga noted that the largest research project of Meta was Infiniband three years ago, but now more than 100,000 GPU clusters have been fully converted to Ethernet. Similarly, the Amazon AWS Large GPU cluster also uses Ethernet completely, indicating that Ethernet is the correct choice.
In addition, for the development trends and innovative breakthroughs of Ethernet, Broadcom has continuously launched two Tomahawk 6 and Jericho 4 products to meet the growing demand for AI computing. First of all, in terms of Tomahawk 6 exchangers, the speed of Tomahawk 6 is twice as fast as the previous generation Tomahawk 5. It is made of telco's advanced 3nm process and has a 100T speed converter. This requires three-layer networks to be used if you want to connect to the same number of XPUs or GPUs. However, using a 100T speed converter requires only one layer of network. In this way, the network structure is not only greatly simplified, but also reduced power consumption and optical component requirements. This also reduces the number of conversion levels and further reduces the delay and improves the efficiency. And these designs are currently being used by customers to build their related systems.
As for Jericho 4 ethernet switch routers, today's super-large AI clusters require about 200 MW of data centers to accommodate. However, many areas are difficult to provide such a large single data center. The Jericho 4 should be born by using an ethernet exchange router that can connect multiple dispersed data centers, up to 100 kilometers, making it look like a large data center.
Ram Velaga emphasizes that the Jericho 4 also uses 3-nanometer process of telectroelectronics, and is also using advanced CoWoS-S packaging technology, containing circuits of more than 1,600 square millimeters (mm²) and combined with HPMS technology, making it one of the most complex chips in the world, and may even be more complex than some GPUs, and its deep-expanding HPMS technology is also amazing.
Ram Velaga emphasizes that the network is the key to realizing distributed computing. Broadcom has sent important information to GPU manufacturers, that is, network and GPU should not be sold in a fixed manner. Broadcom believes that if GPU manufacturers are confident in their products, they should allow the best networks in the market to compete freely, allowing the best networks to emerge, and at the same time allowing the best GPUs to emerge. This open competition model can ensure that users have more choices and promote technological innovation in the entire industry through market forces, rather than limiting competition by closing the environment.
Broadcom further pointed out that Ether has become the first choice for high-performance networks today. Through innovative products such as Tomahawk 6 and Jericho 4, Broadcom not only improves network speed and efficiency, but also solves the expansion challenges of large AI clusters. It also advocates the establishment of a more open and free-competitive Internet ecosystem. In the future, Ether will continue to be the cornerstone of connecting global computing resources and realizing the connection between all things.