What Qualcomm’s rumored exit from data centers means
The tech industry got a jolt last week worse than the 3.5 magnitude quake that hit Oakland, California, on Monday. A report by Bloomberg, citing the usual anonymous sources, said that after a whole lot of R&D and hype, Qualcomm was looking to shut down or sell its Centriq line of ARM-based data center processors.
Qualcomm launched the 48-core Centriq 2400 last November. At the time, potential customers, such as Microsoft, Alibaba and HPE, took to the stage to voice their support and interest.
Qualcomm stops talking about Centriq
Since then, nothing. Qualcomm hasn’t said a word about Centriq on two earnings calls since then. A check with Mercury Research and IDC, which both follow market share numbers pretty closely, found no measurable sales of Centriq products. And none of my usual analysts have heard a word beyond the Bloomberg story.
If you’ve followed any industry, you know that when a company stops talking about a product, it’s probably doomed. The company loses faith before they kill off a product, and silence is a major indicator of it. We’ve heard nothing from them about wins or deployments, and nothing about the next generation.
Cavium launches its own ARM-based processors
By contrast, Cavium, a much smaller player than Qualcomm, last week launched its own ARM-based processors, the ThunderX2, and had all kinds of support for it from Cray, HPE, and Microsoft.
It could just be that Qualcomm, which specializes in low-power, embedded processors for handheld devices, simply decided it didn’t have the spare money and resources for a protracted fight with Intel and the now-revitalized AMD over the data center.
On its most recent earnings call, Qualcomm CEO Steve Mollenkopf told analysts that the is focused on spending reductions in its non-core product areas. Clearly data centers would be non-core.
Qualcomm is still fending off a hostile takeover of Broadcom and is struggling to buy NXP Semiconductors for $44 billion. NXP owns Freescale, which has its own ARM business but mostly on the embedded side, so it’s more synergistic. Plus, Qualcomm authorized the buyback of $10 billion in shares to fend off Broadcom. That’s $10 billion not available for a protracted fight with Intel.
Can anyone compete with the x86
It’s not a good showing for ARM in the data center. Broadcom had a line, called Vulcan, which it sold off to Cavium and now makes up the ThunderX2 line. AMD was working on an ARM chip for the data center called K12 but hasn’t talked about it in years. And Calxeda, the first attempt at a data center chip based on ARM, crashed and burned spectacularly.
The business question then becomes, can anyone compete with the x86 hegemony? Intel owns virtually everything, but there are indications AMD is about to start nibbling at its lunch again. The unpleasant reality is that even if Centriq can offer better price performance than Skylake, it still means a code rewrite, and the server side is the most resistant to change and disruption.
If that is indeed the case, how much traction can Cavium expect to gain? Not much, I suspect. It will probably do well in new deployment scenarios, but as a replacement for Xeon, I don’t think it’s going to happen. Not that they can’t get some decent sales out of it. I just don’t think ARM in the data center is going to be the revolution some have hyped it to be.
Finally, who would buy the Centriq line? Cavium/Marvell would undoubtedly love it. They would have the market all to themselves, and no one can claim monopoly or antitrust issues here. But the Marvell merger might preclude that. A data center monster like Amazon, Google, Facebook, or Microsoft could do it. They all have their own custom chip designs as is and could easily adapt their cloud services to ARM.
If that happens, and it seems more likely, then don’t expect to see those chips on the market. Like Google’s TPU and Facebook’s forthcoming AI processor, these chips will power their data center, but the vendors will keep them for themselves.