Hot on the heels of our overview of the state of the Arm server CPU landscape yesterday, chip maker Qualcomm, which dabbled in Arm server chips a few years back, announced that it was acquiring startup Nuvia for its Arm chips and design team. But don’t get all excited about Qualcomm coming back into the server business, which Nuvia was most definitely targeting since it was founded. Reading the tealeaves in the limited announcement that Qualcomm made, we definitely get the impression that Qualcomm is not interested in designing and selling Arm server CPUs.
Admittedly, this is inference on our part. We have a call logged into Qualcomm’s PR to try to get some insight, but we wouldn’t expect the company to comment anyway with the $1.4 billion acquisition of Nuvia not completed. Qualcomm did not host a call with Wall Street to go over its hopes and dreams with the Nuvia acquisition, and given that Nuvia has said very little about what its Arm server chips would look like anyway, we wouldn’t expect any kind of insight into what Nuvia was doing architecturally to build a better Arm server mousetrap, either.
What we can also infer is that whatever Nuvia co-founders Gerard Williams, Manu Gulati, and John Bruno have cooked up is worth 4.8X the $293 million the company has raised since it was founded in 2019. The second round was $240 million in September 2020, which was a pretty rough time to be trying to raise cash and it didn’t slow Nuvia down one bit.
What Nuvia was promising was a stripped-down CPU aimed specifically at hyperscalers, and without being specific, one that would lead in every key performance metric against which all server CPUs are measured: performance, energy efficiency, compute density, scalability, and total cost of ownership. This engineering problem, Nuvia told us when we profiled the company back in February 2020, was to get the instructions per clock, or IPC, up so Nuvia could deliver “a step function” improvement in performance, and to do so within the same or a lower thermal design point, thus delivering the much-pursued improvement in dollars per unit of performance per watt that drives the buying decisions at hyperscalers. All that Nuvia would say about how it was going to get consistent, double-digit performance increases was that it was going to not use standard Arm Neoverse cores licensed from Arm Holdings, but go all the way and do a clean sheet core and system-on-chip design that was aimed precisely and only at the workloads running at hyperscalers, who can more easily switch CPU architectures can cloud builders, large enterprises, or HPC centers.
If you want to get into all of the big projects that the three Nuvia co-founders have been involved in, and therefore gives them chip cred, take a look at that profile. Williams has a long history of chip design and was the lead architect for the Cortex A8 and Cortex-A15 designs that got Arm chips on the path into tablets and smartphones and moved over to Apple to do many CPU designs. Gulati started out at AMD on X86 processors, did MIPS designs for a while, and then also moved to Apple to design the SoCs that wrapped around the cores designed by Williams. Bruno was a GPU designer at ATI Technologies and then at AMD, and was a system architect at Apple. Both Gulati and Bruno worked a few years at Google designing SoCs and systems for Google’s consumer products.
None of them have a lot of server experience, but they sure do have a lot of client experience. We are not saying they could not build an intriguing and competitive Arm server CPU. But we are saying that Qualcomm, judging from the list of vendors in its acquisition announcement – Microsoft, Google, Samsung, Acer, ASUS, Bosch, Continental, GM, HMD, Honor, HP (not HPE), Lenovo, LG Electronics, LG Mobile, OnePlus, Oppo, Panasonic, Renault, Sharp, Sony, Vivo, and Xiaomi – and the products they hinted about, probably is not going to jump back into Arm CPUs for servers. It spent a lot of energy and time on two generations of Centriq processors – one from 2016 was really a prototype and the other from 2017 was really a testbed for what it could do – before Qualcomm pulled the plug on the whole effort in May 2018. The word on the street was that Google, and possibly also Facebook and Microsoft, were funding the Centriq effort, but none of this was ever confirmed.
What we do know is that Amazon Web Services and now possibly Microsoft have their own, homegrown Arm server design teams, and we think if they do, then a whole bunch of them also do or are tightly aligning with indigenous Arm chip makers in either the United States or China. Those that can make their own, will, and those that can’t, will buy from a company like Ampere Computing and, until now, possibly Nuvia, and maybe someday Nvidia. We think that the big clouds and hyperscalers want control of their hardware, top to bottom, and that means Nuvia found a pretty good buyer without ever having to ship a single chip for a single server.
For its part, Qualcomm can re-enter the server racket if it so chooses with the Nuvia team and use whatever innovations that Nuvia came up with in its Snapdragon client and embedded Arm processors and call it homegrown since no one has ever seen it. And it keeps whatever this Nuvia technology is out of enemy hands – and there are plenty of enemies.