Nvidia officials are taking a significant step in growing the GPU maker’s presence in the rapidly changing data center space and making it a more formidable competitor with Intel with their planned $6.9 billion acquisition of networking vendor Mellanox Technologies.
The company announced the deal March 11, following several days of speculation that Nvidia was interested in buying Mellanox, which had also been getting acquisition attention from Intel, with reports circulating that Intel several months ago had offer $6 billion. Another chip maker, Xilinx, also was among the bidders. The deal is the largest in Nvidia’s history.
Nvidia founder and CEO Jensen Huang said during a conference call to announce the deal that having Mellanox in the fold will increase his company’s capabilities and status in a broad range of data center scenarios, from enterprise and high-performance computing (HPC) environments to cloud and hyperscale infrastructures.
“This acquisition … would be transformative for us, changing us from a chip- and system-level company to a data center-level scale company,” Huang said. “In the future we want to optimize data center-scale workloads, again across the entire stack, from the compute node to networking to storage. For this reason, Mellanox’s system-to-system, data center-scale interconnect technology is important to us.”
The broad interest in Mellanox—which offers a broad networking portfolio, from InfiniBand and Ethernet switches to adapters, silicon and software—comes as data center workloads continue to evolve with the rise of such technologies as artificial intelligence (AI), machine learning and data analytics. Data centers are becoming increasingly data-center and compute-needy, putting greater pressure on architectures for more speed and bandwidth, as well as greater power and cost efficiencies.
Moore’s Law Slows
At the same time, Huang said, Moore’s Law, which has been the driving force behind innovations around CPUs for more than five decades, is beginning to slow. Nvidia a decade ago began offering its GPUs to help accelerate the performance of servers, first in HPC and then in cloud and enterprise data centers, and heterogeneous and accelerated environments will be key to addressing these more modern workloads, he said.
With the growth of AI, analytics and other data- and compute-intensive workloads, organizations will need tens of thousands of server nodes working as a single unit. The network will be the connecting fabric that ties the environment together, said Huang, who several years ago put AI and machine learning at the center of Nvidia’s growth plans.
“CPU advancement is slowing just as computing demand is skyrocketing,” he said. “As a result, accelerated computing and accelerated networking … is a fundamental path forward. … The dynamic that is happening here is that in the future, it won’t be just server-scale computing that people do, but it will be data center-scale computing, where the network becomes an extension of the computing fabric.”
The combination of Nvidia and Mellanox will make for a formidable data center company, according to Patrick Moorhead, principal analyst with Moor Insights and Strategy.
“This deal makes Nvidia much more of a data center player, particularly where the market is headed, and that is more heterogeneous computing,” Moorhead told eWEEK. “The more heterogeneous computing is required, the [more] higher performance networking is required.”
Nvidia Addresses Performance Issues
Rob Enderle, principal analyst with the Enderle Group, told eWEEK that the deal “allows [Nvidia] to address directly some performance issues with their HPC solutions that, up until now, were outside of Nvidia’s product set. Nvidia solutions tend to generate massive data streams because of their visual nature and they likely often bottlenecked at the Interconnect layer. Mellanox’s technology already addresses this problem, but a closer collaboration between the firms should have a significant benefit from focused tuning that should uniquely benefit Nvidia.
“Right now Nvidia is more of a parts vendor than a solutions vendor, but, as the need for their technology in the data center has expanded, they’ve been forced to deal with other parts of the ecosystem, in this case interconnect technology,” Enderle said. “This will strengthen their portfolio and you should expect them to make additional acquisitions as they discover new system bottlenecks adversely impacting the performance of their offerings.”
Nvidia and Mellanox already have a long history of collaboration, with Huang noting that their technologies are in more than half of the systems listed on the Top500 list of the world’s fastest supercomputers, including Summit and Sierra, ranked No. 1 and 2 on the latest list released in November. Products from both also have been adopted by the top system OEMs, including Dell EMC, Hewlett Packard Enterprise, IBM, Lenovo and Cisco Systems, and cloud service providers like Amazon Web Services, Microsoft Azure, Google Cloud Platform, Oracle Cloud and Alibaba Cloud.
At the same time, Nvidia has leveraged Mellanox technologies in its own systems, including the DGX-2 server (pictured) for AI workloads, which uses Mellanox InfiniBand and Ethernet technologies.
Massive Addressable Market
Nvidia officials said a combined company will address a market that includes compute and high-speed networking that will be more than $60 billion by 2023.
The deal has been approved by the boards of directors of both companies, and will subject to the approvals of Mellanox shareholders as well as regulatory agencies. Nvidia officials said they don’t see a lot of overlap in what Nvidia and Mellanox offer and that the combined company will continue to invest in employees and operations in Israel, where Mellanox is based. The deal is expected to close later this year.
The changing dynamics and evolving workloads in modern data centers also are pushing other chip makers to adapt. Intel, which continues to hold a dominant portion of the data center chip market—north of 95 percent by many estimates—not only continues to innovate on its Xeon CPUs but also now offers accelerators like field-programmable gate arrays (FPGAs)—thanks to its $16.7 billion acquisition of Altera in 2015—and ASICs via its purchase of eASIC last year, and is planning to introduce its own discrete GPUs in by 2020.
Advanced Micro Devices is readying the next generation of its Epyc server chips based on its Zen microarchitecture, which are due out this year, and in 2018 unveiled its latest Radeon Instinct GPU accelerators. Xilinx, which made its name as an FPGA maker, now is offering compute acceleration platforms such as its Versal offering for AI workloads, which was introduced in October 2018. The company also put its FPGAs into accelerator cards for the data center.