
As one year ends and another begins, this is often the time when people change jobs and companies change strategies. So it is never much of a surprise to see key people in datacenter compute change jobs, seeking new opportunities and better strategies, or to hear about potential acquisition targets.
Oddly, the scuttlebutt and job changes are about datacenter CPUs when what the world is focused on – and what an increasing amount of datacenter dollars are spent on – are datacenter GPUs and the interconnects that lash them together.
The big move this week, which is not a rumor but a fact, is that Intel Fellow and chip architect Sailesh Kottapalli, which was at the chip maker for nearly three decades and who was the main designer of the last several generations of Xeon datacenter CPUs, has joined Qualcomm as a senior vice president. Kottapalli did not give out his precise job description in his job change announcement on LinkedIn, but it is not hard to figure that he will be working on datacenter compute engines of some kind.
The circumstantial evidence that Qualcomm is going to take a third run at the datacenter CPU market is mounting, particularly with a December 9 job posting by the Qualcomm Data Center team saying that the company “is developing High performance, Energy efficient server solution for data center applications.” Not sure why High and Energy are capitalized in that description, but perhaps that is a subliminal message to job seekers to drink a lot of coffee and be motivated. In any event, that particular job is for an SoC security architect, and is obviously not the job that Kottapalli is taking. But it seems unequivocal that Qualcomm is rubbing its hands together when looking at the datacenter and a particularly weakened Intel with its underbelly exposed as it wrestles with foundry and datacenter product line issues.
Qualcomm released a prototype Arm server CPU back in 2016 under the guiding hand of Anand Chandrasekher, the general manager at Qualcomm Datacenter Technologies and formerly one of the top execs at Intel in charge of workstation and then mobile products. The “Amberwing” Centriq 2400 processor, with 48 Armv8 cores, was unveiled in November 2017, and by May 2018 Qualcomm pulled the plug on its server CPU aspirations. This was before the hyperscalers and cloud builders excepting Amazon Web Services started designing their own Arm server CPUs, and the rumor was that Google – and possibly Facebook and Microsoft –funded the Centriq effort. This was never confirmed, but perhaps the Amberwing chip underwhelmed and the price was too high.
Since then, of course, Google and Microsoft have created their own respective Axion and Cobalt Arm server CPUs chips, AWS has deployed millions of its own Gravitons, and other hyperscalers and cloud builders are working on their own CPU designs, too. This all begs the question as to why Qualcomm thinks it can break into the server CPU racket even with a substantially weakened Intel, albeit one with a reasonable X86 product line even if it cannot beat AMD Epycs at its own X86 game.
When Qualcomm acquired Arm server CPU startup Nuvia in January 2021 for $1.4 billion – a company that a year earlier had told us it was definitely creating server CPUs – Qualcomm put the kibosh on the idea that it wanted to resurrect its server CPU aspirations. Qualcomm did create its own Cloud AI 100 matrix math accelerators for AI inference, which we profiled when Cerebras Systems added them as sidecars to its third generation waferscale WSE3 matrix engines for AI training.
With the major hyperscalers and cloud builders representing more than half of server CPU shipments and slightly less than half of the server revenues in the worldwide systems market, and with these same tech titans all having their own Arm server CPU designs and compelling economic reasons to use their homegrown Arm server CPU chips, it would seem that now is not a good time to launch a third Arm server CPU effort (if you include Nuvia) at Qualcomm. Then again, more and more applications are written for Linux and that means they can run on Arm and are not dependent on X86 servers. So there might just be a place for an Arm server CPU maker in the market to address the needs of enterprises and service providers who are not big enough to design their own chips.
SoftBank To Buy Ampere Computing?
It would seem that Ampere Computing, which was founded by a bunch of ex-Intel execs and that had its initial Arm foundation from the team and intellectual property of Applied Micro, one of the early Arm server chip designers, should be that second source. It has put multiple generations of its Altra processors as well as its 192-core “Polaris” AmpereOne into the field, after all, and is working towards getting a 256-core “Magnetrix” out this year and a 512-core “Aurora” chip out further down the road.
Last week, Bloomberg reported a rumor that Japanese conglomerate SoftBank, the majority stakeholder in Arm Ltd, was looking to buy Ampere Computing, which has taken investments not only from Carlyle Group but also from Oracle (which has a 29 percent stake) and SoftBank (whose stake is unknown). This is an odd bit of rumor, considering that Ampere Computing no doubt wants to go public. But with 1,110 employees and hyperscalers and cloud builders focused on their own Arm server CPU efforts, perhaps Ampere Computing is stretched thin. Or perhaps Carlyle Group and/or Oracle want to cash out for their own reasons.
And perhaps SoftBank really does want a working CPU from Ampere Computing and a working AI accelerator from GraphCore, which it acquired in July 2024, to make its own systems and install them in its own datacenters, which it then uses and also rents out. That’s what the big clouds are doing, and as long as the AI models run, a lot of users won’t care what they run on so long as the price is reasonable.
We have a hard time believing that SoftBank can put together a company, maybe using the Ampere Computing brand, that will sell iron into enterprises, governments, and lower-tier service providers, but we do think if it created infrastructure that abstracted the hardware away and just offered CPU cycles and accelerator tensor cores running open source models like Llama from Meta Platforms, it could seed lots of clouds that could do AI crunching as a service – and possibly undercut the pricing of the big clouds and still make a profit. This is no different, in fact, than what AWS, Microsoft, and Google are doing with homegrown CPUs and AI accelerators compared to running AI on X86 servers with Nvidia GPUs.
From Intel To Microsoft To Google
People change jobs for a lot of reasons, but sometimes it is because they are a hot commodity. This appears to be the case for Rehan Sheikh, who has been named vice president of global silicon chip technology and manufacturing at Google as of December 2024.
Sheikh started out his engineering career as a member of the technical staff at AT&T Bell Labs back in the early 1990s, and spent 24 years at Intel, between 1997 and 2021, becoming one of its experts in high volume manufacturing, power and performance tuning, quality and yield control, and post-silicon engineering. These skills made him an Intel Fellow, and that skillset was valuable to Microsoft when it started its own Cobalt Arm server CPU and Maia AI accelerator efforts several years back. Microsoft made Sheikh a general manager of its silicon manufacturing and packaging engineering, and a little more than a year ago raised him up to a vice president.
It looks like Google Axion and TPU v6 have a new shepherd making sure they get to the datacenter on time and with good yield.
Be the first to comment