Generative AI drives data center spending and 800 gigabit Ethernet
The surge of interest in Generative AI based on Large Language Models (LLMs) is driving the need for a large increase in processing power and consequently energy consumption.
Generative AI is already being used to generate text, software code, and images, with video content generation likely the next big application. Other applications such as the metaverse, games using embedded Generative AI for characters, and teaching are also of immense interest to developers and consumers.
NVIDIA ACE for Games is a good example of how Generative AI is attempting to revolutionise how games are conceived, produced, and played. The creation of non-playable characters (NPCs) has evolved as games have become more sophisticated. However, the number of pre-recorded lines has grown, the number of options a player has to interact with NPCs has increased, but these player interactions still tend to be transactional, scripted, and short-lived. According to NVIDIA Generative AI can make NPCs more intelligent by improving their conversational skills, creating persistent personalities that evolve over time, and enabling dynamic responses that are unique to the player.
Demand for Generative AI has grown rapidly but its use is set rise exponentially going forward. However, the cost of delivering the processing power and energy required to drive this revolution are going to be a limiting factor in spite of improvements in algorithm and hardware effciency. According to two recent reports this rising demand is driving data center growth and is set to result in a surge in 800 gigabit Ethernet switch adoption.
According to a recently published report from Dell’Oro Group, worldwide data center capex is forecast for a compound annual growth rate of 18 percent as investments shift towards AI. The group anticipates that accelerated computing that are optimized AI workloads will account for a quarter of the data center capex.
“Accelerated computing optimized for domain specific workloads such as AI is forecast to exceed $200 billion by 2028, with the majority of the investments deployed by the hyperscale cloud service providers,” said Baron Fung, Senior Research Director at Dell’Oro Group. “In order to drive long-term sustainable growth, the cloud service providers will seek to streamline general-purpose computing infrastructure costs by transitioning to next-generation server platforms and rack-scale architectures. We also anticipate increased vertical integration efforts by the hyperscalers to control costs and bring further optimizations for their full stack. Meanwhile, the enterprise segment faces near-term headwinds related to economic uncertainties, and will adopt a hybrid cloud model for AI and traditional IT workloads,” explained Fung.
Further, Crehan Research predicts that 800 gigabit Ethernet (GbE) switches will see rapid customer adoption and will surpass 20 million ports in annual shipments within four years — equating to the fastest ever data center Ethernet switch speed ramp. The key driver of the fast adoption is generative AI, which requires much higher networking bandwidth than other workloads.
“Generative AI is potentially the biggest killer application to emerge in decades,” said Seamus Crehan, president of Crehan Research. “As the de facto data center networking technology, with an installed base of over 400 million switch ports as well as the proven ability to incorporate additional features to handle the requirements of new applications, we expect Ethernet to be the main enabler of these workloads,” he said. “Already, Ethernet networking is seeing generative AI enabling deployments in five of the top seven hyperscale cloud service providers, as well as the world’s fastest supercomputers including the Frontier exascale system and the Aurora system.”
In conjunction with strong adoption of high-speed Ethernet switches to handle generative AI applications, Crehan’s report also predicts rapid growth in 100-Gbps SerDes shipments. The company expects this technology to surpass 50-Gbps SerDes shipments to become the most-deployed data center Ethernet switch SerDes speed within three years.
“100-Gbps SerDes are the building blocks for the next generation of ultra-high-bandwidth data center Ethernet switch silicon products,” Crehan said. “That includes Broadcom’s Tomahawk 5 and Jericho3-AI as well as Cisco’s Silicon One G200, which will serve as the data center networking foundations for the next phase of generative AI adoption.”
www.delloro.com
www.CrehanResearch.com
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
