Intel’s New Nemesis? Qualcomm Shows Off Its 10 nm Server Processor

The 10-nanometer server processor market officially opens its gates in the second half of next year, and today, we’ve seen the first physical evidence that Intel will not be taking the field alone.  In the clearest sign to date that it could, at the very least, meet Intel day-and-date, on Monday, Qualcomm staged a demonstration to select participants of its first ARMv8-based server processor: the Centriq 2400, formerly code-named “Amberwing.”

We’re hearing the first news of this demonstration, in a Qualcomm announcement Wednesday.

Qualcomm is calling this first chip in the series “purpose-built for performance-oriented datacenter applications,” in promoting today’s news to the press.

A Qualcomm spokesperson provided to Data Center Knowledge on Wednesday as much information as the company is willing to share with the public at this time, about the demo.  It was an analytics application involving both Apache Spark and an unbranded Hadoop distribution, along with the APIs of Twitter and Google.  Java middleware was present, and an unnamed Linux provided the basic platform.  No virtualization or containerization was involved, and all software used in the demo was open source.

Qualcomm-Centriq-REP_full-lg-640-px

A third party was contracted to produce the servers used in the demonstration, which did not carry major brand names.  Although the chip design is said to support as many as 48 cores, the quantity of cores used in the demonstration is not being shared publicly.  However, Qualcomm did share with Data Center Knowledge the first pictures of the 1U server platform involved in the demonstration, which did carry the Qualcomm Centriq trademark, along with the designation “QDF2400 REP.”

Qualcomm-Centriq-REP-Angle-lg-640-px

No performance indicators are currently available, and it’s telling that Qualcomm did not choose to use an industry standard benchmark for this demo.  The company may be in a greater position to provide more deterministic performance data at some point during Q1 2017, we were told.

The spokesperson confirmed that Qualcomm has begun sampling servers running on Centriq processors with select customers.

Finally Putting 3D to Work

Centriq is being fabricated using the FinFET process, which is Qualcomm’s choice for implementing 3D transistors.  One way of overcoming the physical limitations of shuttling electrons in close proximity on a processor, is by effectively “printing” that processor in layers, creating 3D components.  While Intel and Micron Technology have been collaborating on an architecture called 3D XPoint, its use at present is limited to memory modules.

With FinFET, the planar substrate acts as a lowest-level foundation layer, letting the gate of the transistor wrap around the source and the drain.  This has the immediate benefit of radically reducing leakage, enabling lower voltages.  It’s these low voltages that are absolutely necessary for implementing a lithography process as tiny as 10 nm, and FinFET is one way to make this happen.

While the concept is already a quarter-century old, it’s taken this long for an evolutionary event — in this case, the impending dead-end for Moore’s Law in its current form — to compel chip manufacturers to give it a shot.

Qualcomm declined comment on the details of its FinFET process at this time.

An Athlete in a Different Sport

The man driving Qualcomm’s production of Centriq is Anand Chandrasekher [pictured top], who may be best known for having led Intel down a path toward a dead-end: as the head of Intel’s ill-fated Ultra-Mobile PC project with Microsoft a decade ago.  Qualcomm hired him away from Intel in 2012, at first appointing him Chief Marketing Officer.

After a series of incidents in which he publicly called Apple’s A7 system-on-a-chip a “gimmick,” and characterized the idea of an 8-core, 64-bit processor as “dumb,” the company took the unusual step of publicly censuring Chandrasekher, reassigning him to a vice presidency.

The following year, in a public conversation with All Things D’s Ina Fried, Chandrasekher chastised Intel for acting like a prize athlete competing in the wrong sport — in this case, mobile devices, where its ambitions were crushed by Apple.  Let a company built for mobility compete in mobility, he argued at the time.

Ironically, three years later, he’s the one driving Qualcomm down its own unfamiliar path, touting the benefits of a processor with as many as 48 cores.

“This migration to a cloud model has. . . created a major power shift in the supply chain,” wrote Chandrasekher, in a Qualcomm company blog post Wednesday, “with mega data centers sourcing server platforms directly from ODMs. . . Most data center software is based on open source software projects.  This enables Qualcomm, as well as other ARM ecosystem partners, to work within these open source projects to implement support for ARMv8.”

During his Monday presentation, Chandrasekher made the case that the PC had been the driving factor behind the process technology of processors.  But as the PC sales slump became a permanent trend, mobile devices took over that role.  Now mobile phones, he argued, are the principal drivers behind processor fabrication, thus giving Qualcomm an opportunity to take the lead in the industry.

A Friendly Dragon?

Last month, Qualcomm’s Falkor processor core (named for a furry, benevolent dragon in the movie, “The Neverending Story”) was officially inducted as a supported processor by the community supporting the LLVM systems compiler project, as indicated by this update page on GitHub.  This means developers building low-level software, such as operating systems and libraries, may use LLVM to compile C and C++ code for ARMv8 chips using the Falkor core, Centriq 2400 being the first series to do so.

A recently uncovered, “leaked” copy of Intel’s processor roadmap, appears to indicate that the first series of that company’s 10 nm processors, code-named “Cannonlake,” is due for production very late next year — on the tail end of Q4.  If Qualcomm makes good on the front end of its H2 2017 promise for Centriq, it could beat Intel to market by as much as five months.