Inside Lenovo’s AI Factory
The company’s innovation center in Budapest offers customers the opportunity to run proofs of concept Torrential rain is not the ideal weather to launch 5,072 solar panels onto a factory roof. However, on a meteorologically disastrous day in April, DCD braved the slippery conditions of a somewhat scary outdoor spiral staircase to see the culmination […]
The company’s innovation center in Budapest offers customers the opportunity to run proofs of concept
Torrential rain is not the ideal weather to launch 5,072 solar panels onto a factory roof. However, on a meteorologically disastrous day in April, DCD braved the slippery conditions of a somewhat scary outdoor spiral staircase to see the culmination of the latest development at Lenovo’s Budapest, Hungary, facility. In 2020, the land where the factory now stands was a cornfield. At the height of the COVID-19 pandemic, the building’s outer shell was built in 10 months, and it took another seven months to obtain all the relevant licenses and build the production lines. The factory now has a total floor space of 49,000 square meters (527,400 square feet) and ships between 600,000 and 800,000 Lenovo products, including commercial and enterprise servers and storage, high-performance computing (HPC) solutions, and water-cooled servers. Lenovo ships these products to 2,500 customers in 70 countries each year. From day one, the factory was designed with sustainability in mind. In the winter months, cold air is pumped in from outside to keep the factory cool, while heat generated by the servers is used to heat water that circulates through the building’s radiators. The newly introduced solar panels provide 3MW of solar power capacity, which, while not enough to power the entire factory, does provide enough energy to run its HPC and AI Innovation Center, which is located at the heart of the operation.
The Innovation Center provides Lenovo ThinkSystem and ThinkAgile hardware and solutions to support demonstrations, proof of concept and benchmarking for the company’s EMEA customers, and each Lenovo ThinkSystem SR680a V3 and ThinkSystem SR780a V3 can support eight Nvidia H100 and H200 Tensor Core GPUs.
The center also offers remote VPN access to customers and partners, as well as providing Lenovo’s first liquid cooling demonstration and testing center in Europe. Likewise, the company’s global HPC and AI testing facility is supported by the Lennox HPC Cluster architecture, which again provides solution benchmarking, proof of concept, and demonstration capabilities to the company’s global customer base.
A second AI innovation center to serve North American customers is located at Lenovo’s “mega sites” facility in Monterrey, Nuevo León, Mexico. Speaking to DCD in London ahead of the factory tour, Scott Tease, who leads Lenovo’s supercomputing, high-performance computing and AI systems team, says one of the biggest challenges data center vendors face in the AI era is power constraints, as many data centers aren’t currently equipped to handle the kind of density these workloads require.
“Most of our data centers are designed for 8 to 15 kW per rack, but one of these generative AI servers is 10 kW, just one. “Imagine having a rack that’s 40 rack units high with a single server in it. It must feel like you’ve wasted some of the most expensive data center space in the world,” he says. More powerful hardware means companies can now run workloads that previously required 42 servers across 15 or 18. But Tease isn’t sure that the increased efficiency justifies the power these servers now consume, a problem he says is best illustrated by the proliferation of GPUs.
“With traditional CPU servers, the most you could get would be maybe a 40kW rack,” Tease says. “It would really be going beyond that, really doing something really amazing to get to that density.”
“With GPU systems, we’re already shipping racks of GPU servers with 80 to 85 kW of power per rack.”
Tease notes that because of supply chain constraints, GPUs are now being sourced from vendors other than Nvidia. It now has competition in the form of AMD’s MI300X and Intel’s Gaudi3, both of which will begin shipping later this year.
“[AMD and Intel offer] a very compelling accelerator for these large systems,” he says. That’s one of the benefits of the AI Innovation Center. In partnership with AMD, Intel, and Nvidia, customers can test out new data center hardware and solutions before deciding whether to invest a significant amount of money or join a long waiting list. Built to perfection
Back in Budapest, speaking to DCD on the day of the solar panel unveiling, Brian Jaeger, director of supply chain services at Lenovo, explained that the AI Innovation Center has the capacity to support about 20 test racks of concept (POCs), meaning the number of customers who could use the center’s capabilities at any given time could vary depending on the number of servers your POC requires.
“Typically, a proof of concept lasts between 30 and 90 days,” he says. “That’s how long we try to work with.”
Lenovo plans to increase that capacity over the next year, but hasn’t said how much. Jaeger added that one of the benefits of having the Innovation Center located within the factory is that it allows the company to offer nearly its entire product portfolio for inclusion in proofs of concept, because all of Lenovo’s supply chain materials are located in the same facility.
“We can basically build the exact configuration that a customer is looking for,” he says, something that wouldn’t be possible if it were located in a Lenovo development center, as access to inventory would be much more limited. “We can build exactly what customers think they want and basically offer a ‘try before you buy’ environment,” Jaeger adds.
This also allows them to build configurations using all the necessary components, rather than testing a configuration that uses the customer’s chosen AI accelerator but a basic version of a storage solution. The Innovation Center was built specifically to support AI workloads, so it’s no surprise that Jaeger says there’s a lot of focus on AI right now and it’s the space where Lenovo customers are seeing the most gains.
“Our focus here over the last 12 to 24 months has been to continue to develop and expand our AI portfolio,” he says, adding that the data center market has seen a number of trends over the last 20 years before, more recently, focusing on AI. Despite some recent skepticism about the long-term promise of generative AI in particular (a Goldman Sachs report earlier this year said that large companies will spend $1 trillion on AI hardware over the next few years without a clear plan to get a return on their investment), Jaeger believes the AI hype will last for some time.
“We continue to expand our portfolio to support it, and we’re making significant investments in these high-end AI systems that are coming to market, so I absolutely believe [AI] is here to stay,” he says. “Now, will it change direction in 10 years? Probably. But for several years, I think we will continue to focus on developing AI capabilities.”
Was this article useful to you?
0 Feedbacks