AI hardware demands drive biggest server, chip deals in 2019

HPE gains supercomputing clout through Cray

Heading the list was the out-of-the-blue $1.3 billion acquisition of long-time supercomputer maker, Cray, Inc., by HPE. With this deal, HPE hopes to revive the fortunes of the supercomputer market by cascading the technology down into commercial enterprises where IT shops can harness its power to drive AI applications. HPE's CEO Antonio Neri notes the deal brings Cray's "foundational technologies" on which to build the next generation of server and communication of products.

The deal, expected to close in 2020, could lend HPE more credibility at the high end of the market following the disappointment of The Machine -- a server chock full of next generation technologies. HPE introduced the system over four years ago but it has still not found its way to market.

The Cray deal also opens the door for HPE to compete for lucrative government contracts up for bid, including the Frontier supercomputer at the U.S. Department of Energy's Oak Ridge National Laboratory that requires systems with exascale capabilities. This is an area where Cray has had much greater success than HPE.

IBM quantum system makes its debut

After 30 years of research and development, IBM finally delivered the industry's first integrated quantum computer, the 20-qubit IBM Q System One. While Big Blue delivered the complete hardware-software-communications package, including an SDK for corporate developers and a 9-foot airtight glass-paneled cube to house the system, it will be some time before IT personnel have the necessary skills to deliver exploitive commercial applications.

That lack of experience, coupled with the high cost of maintaining the system on-premises, means IBM will deliver quantum capabilities to customers via cloud-based services. IBM officials expect to make a 50-qubit system available to developers later this year, which should allow developers to create more meaningful applications.

"Initially, what you will see is more apps for things like managing massive sensor environments and data sets where the scope and complexity is huge," said Judith Hurwitz, president and CEO of Hurwitz and Associates LLC, a research and consulting firm in Needham, Mass.

IBM also plans to open the IBM Q Quantum Computation Center later this year in Poughkeepsie, N.Y. The center will be available to corporate developers and business partners that join the IBM Q Network.

Since the debut of the IBM Q System One, a handful of competitors have shown off or talked about the progress made with their systems, most notably Google, Microsoft and Rigetti, although only Rigetti has made its platform accessible to external users or developers.

A handful of smaller startups showed off their quantum computing systems in early 2019, including IonQ. The company hopes to deliver a system using atoms and lasers to store and access information, to be used in IT shops at room temperature, eliminating the need for costly climate-controlled environments.

AI hardware deals

In another significant acquisition, NVIDIA purchased Mellanox for $6.9 billion in hopes of strengthening each company's position in the increasingly competitive -- you guessed it -- AI hardware market.

The deal is as much a defensive move as it is an offensive one, given the escalating competition NVIDIA's GPU technology faces from AWS' Inferentia chip and from the other direction, Cray's Slingshot -- an interconnect aimed at HPC and AI platforms. Together, the combined company figures to be stronger together against such competitors and a slew of smaller competitors entering the AI hardware market.

The two companies aren't strangers to each other. Both contributed technology to Sierra, the Department of Energy's (DOE) supercomputer and one of the fastest systems in the world. Sierra uses NVIDIA's GPUs and Mellanox Infiniband interconnects, along with IBM's Power 9 chips. Both companies also contributed the same core technologies to the DOE's Summit supercomputer, which now tops of list among the world's 500 fastest computers.

Moore's Law winds down

Another notable development – more akin to an ongoing evolution – is the winding down of Moore's Law. Named after Gordon Moore, former CEO of Intel, the thesis states the number of transistors in an integrated circuit doubles every two years, thereby adding steady speed and performance to a CPU.

But chip companies are running out of real estate to cram in more transistors. This has given rise to a wide variety of support chips, most notably GPUs, and other board-level sub-systems that are rapidly becoming standard componentry in server hardware. All of this aimed at churning AI and machine learning workloads faster.

Data centers can no longer carry out their compute on general purpose CPUs and are adding GPUs and FPGAs to get the work done.

"Data centers can no longer carry out their compute on general purpose CPUs and are adding GPUs and FPGAs to get the work done," said Peter Rutten, research manager for IDC's Enterprise Infrastructure Practice.

A report Rutten released earlier this year titled "Worldwide Accelerated Server Infrastructure Revenue 2018 – 2022" forecasts that revenues in this segment will grow from $10,257.4 billion in 2018 to $25,555 billion by 2022.

Even mainframes are facing their own version of Moore's Law, as evidenced by IBM's z Series. Over the past decade, Big Blue has invested heavily in specialty chips to boost its power to run AI and machine learning workloads to keep the venerable server relevant in today's market, Rutten said.

"It is getting harder (for IBM) to deliver more performance as they run into the wall of Moore's Law," said Mike Chuba, vice president of infrastructure and operations at Gartner. "With the next two releases (of their mainframe), it will be the same 8% to 10% performance improvements. The days of 25% to 30% performance improvements are nearing an end," he said.

Amadeus IT Group, a travel software company, recently hooked up with engineers from ETH Research to study the use of GPUs and FPGAs in inference applications based on machine learning. The research showed that GPUs tend to be users' first choice to solve their latency problems involving AI applications, but FPGAs require significantly less power, making them a more attractive candidate for AI algorithms.