In the 20 or so months since ChatGPT began sparking enormous interest in generative AI, many companies continue to search for ways to deploy the technology. When immense data volumes are involved, the hurdles in leveraging AI can seem daunting, according to Lenovo's Robert Daigle. Often, large language models and a multitude of GPUs must be deployed at great expense. But that is just one scenario. Robert says some companies need help just figuring out where to start.
Though Lenovo is famous across the globe as a juggernaut personal and business systems manufacturer, the multinational company also provides AI solutions and services to businesses operating in a wide range of markets. Robert oversees Lenovo's global AI business, the group tasked with assisting clients on setting the right goals, overcoming complexities, and building infrastructures that drive the most value from AI.
"We can be very prescriptive then on right sizing the whole solution — not just the hardware, software, [and] services," Robert said. "We want to right size the whole solution, so we're not over prescribing more than what a customer needs, because we don't want to create artificial barriers to get started…if we can help accelerate time to value, help customers get started really quickly and see value from it, they're going to continue the investment."
AI adoption in the enterprise has lagged due to the mistaken perception that every AI solution requires large language models and massive investment, Robert said. He noted that even within individual organizations, business challenges are diverse, and the best solution doesn't always require deploying the most powerful AI.
"You don't need a GPU for everything," Robert said. "You don't need the highest performance interconnected accelerators for every use case. That's great if you're trying to train a hundred-billion-plus parameter model from scratch. Yeah, having that really high-end compute with interconnected GPU's in a system, it's the right choice. But for a lot of enterprise customers, they're just going to be looking at inferencing and maybe doing some fine tuning… and there's even a trend towards smaller language models."
For smaller language models and inferencing, CPUs and lighter weight accelerators are more than adequate, Robert said. That's why one of the keys to Lenovo's "AI for All" strategy is providing customers with more options and flexibility. One way that Lenovo supplies companies with greater choice is by providing a wide range of hardware and software.
In April, Lenovo introduced a suite of servers and HCI devices, including the ThinkSystem SR685a V3, that support hybrid AI-centric infrastructure systems and "compute intensive workloads." ThinkSystem SR685a V3 is powered by 4th generation AMD EPYC processors as well as AMD's Instinct MI300X GPUs and other GPU options. The system is designed as a solution for on-prem as well as for public AI cloud services.
Lenovo also helps address concerns involving security, eliminating bias from AI and protecting the environment. To assist customers in building responsible and ethical AI, Lenovo shares a multi-step review process initially created for internal use.
"If it has AI in it in any form or function, even if it's just something we're using internally, it has to go through this review," Robert said. "We're looking at how the model is developed. How was the data source that it was trained on? We're looking at not just security things, but also diversity and inclusion aspects. Are there biases in the data? Are there security vulnerabilities that are introduced by this AI solution?"
When it comes to energy consumption, Robert said that companies rightfully fret about power consumption. He marveled that some data centers are gobbling up as much energy as small countries. He said that for companies employing LLMs, solutions include using energy efficient processors, such as AMD EPYC CPUs, and potentially liquid cooling.
Direct to node warm-water cooling solutions for data centers have impressed Robert. He said that this system can extract 90% of the heat. While use of warm water may sound counterintuitive, the method offers more efficiency than cold water or some glycol based solutions, he said. But Robert advised that even before managers consider which cooling system to employ, they should ask themselves two fundamental questions:
- Do you need a large language model?
- Do your computational needs require GPUs?
He predicted that the answer to these questions for many companies, especially as business leaders begin to understand the strengths and limitations of AI, will be "no."
