At Computex, NVIDIA announced NVLink Fusion to enable “semi-custom” AI data center compute architectures; Fujitsu and Qualcomm named as launch CPU partners
The world of AI infrastructure, particularly the underlying silicon that powers so-called “AI factories,” is an increasingly competitive space. NVIDIA obviously holds pole position with its high-powered (and expensive) GPUs which are instrumental in training advanced large language and multi-modal models. But there’s seemingly a renewed interest in CPUs as the center of AI gravity gradually transitions from training to inference.
This week at Computex in Taiwan, NVIDIA announced NVLink Fusion, new silicon that leverages the NVLink computing fabric to enable integration of third-party CPUs with NVIDIA’s GPUs to make what the company referred to as “semi-custom AI infrastructure.” Before we get into the launch partners, let’s first contemplate why NVIDIA would do this at all given that it also sells CPUs and, indeed, rack-scale solutions that would let a user stand up an AI factory using only NVIDIA products.
NVIDIA CEO Jensen Huang, in a statement, called out how NVLink Fusion allows for construction of “specialized AI infrastructures” which will be necessary as enterprises of all sizes and types adopt AI. If the total addressable market is all companies and workflows, there necessarily won’t be a one-size-fits-all solution. And this is the durable narrative from Huang and many other tech leaders. “A tectonic shift is underway: for the first time in decades, data centers must be fundamentally rearchitected — AI is being fused into every computing platform,” he said.
Sure. But there are strategic reasons beyond specialization that, when considered as a whole, represent a strategic evolution. By facilitating interconnect between its GPUs and third-party CPUs, NVIDIA can drive its tech into new sectors that were perhaps previously inaccessible due to any number of constraints or customer preferences. It also meets customers where they are; the world of enterprise compute hardware is diverse, and opening up interconnect creates an entry point to marry NVIDIA GPUs with existing infrastructure.
Further, the net-net for NVIDIA could be increased demand for its GPUs given a new degree of versatility. And casting a wider collaborative net sets the stage for more co-development, more innovation, and more sales. Finally, NVIDIA knows quite well that competitive dynamics can change fast. By supporting integration of its GPUs with other firms’ CPUs, it’s less likely customers could jump to entirely different platforms.
With the announcement of NVLink Fusion, NVIDIA named Fujitsu and Qualcomm as CPU partners who can “couple their custom CPUs with NVIDIA GPUs in a rack-scale architecture to boost AI performance.” The inclusion of Qualcomm in the launch here is notable. Last week, as US tech leaders traveled to the Middle East to shake hands on a number of AI and AI-related deals, Qualcomm quietly announced it was re-entering the data center CPU market. Specifically, Qualcomm is working with Saudi Arabia’s HUMAIN, which itself is also working with NVIDIA to build out a massive AI infrastructure plant, on what Qualcomm CFO/COO Akash Palkhiwala described as “data center solutions, both for inference and for CPU chips.”
Speaking at a JP Morgan conference, Palkhiwala talked through data center CPUs as a new frontier for Qualcomm’s diversification strategy, which to date includes automotive, industrial IoT, PCs and XR. “The change in the data center that’s happening is obviously the move to inference…and the importance of low power,” he said. “And that’s where Qualcomm shines…The announcement we made yesterday [May 13]…is really us bringing those technologies to the data center.”
Qualcomm CEO Cristiano Amon teased the expansion into data center CPUs at the very end of his Computex keynote. He provided a bit more color in an interview from Taipei with CNBC. “I think we see a lot of growth happening in this space for decades to come, and we have some technology that can add real value…I think we have a very disruptive CPU…As long as…we can build a great product, we can bring innovation, and we can add value with some disruptive technology, there’s going to be room for Qualcomm, especially in the data center.”
Amon, as well as Palkhiwala, both referenced Qualcomm’s ability to deliver high-performance, low-power compute for AI. This is important. There are trillions of dollars being spent on AI infrastructure and a major constraint is around power. And not just access to power, but the opex associated with that power. If you can manage that particular line item and still deliver the goods — performant AI solutions — you’re onto something. As Amon said in his keynote, “We have some very interesting IP on CPU.”
Qualcomm’s custom Oryon CPUs, are present in its automotive, mobile and PC platforms. Oryon is based off of the company’s 2021 acquisition of Nuvia. In the mid-2010s, Qualcomm announced a data center chip program that resulted in the Centriq 2400 series for server OEMs.
NVIDIA also announced Fujitsu as a CPU partner. Fujitsu CTO Vivek Mahajan said the company’s 2-nanometer MONAKA CPU combined with NVIDIA’s full-stack “delivers new levels of performance…Directly connecting our technologies to NVIDIA’s architecture marks a monumental step forward in our vision to drive the evolution of AI through world-leading computing technologies — paving the way for a new class of scalable, sovereign and sustainable AI systems.”
