This market research report was originally published at Yole Développement’s website. It is reprinted here with the permission of Yole Développement.
Even setting aside the roller-coaster of COVID-19, 2020 brought changes to the traditional processor landscape, announces Yole Développement (Yole), in its Processor Quarterly Market Monitor, Q1 2021. Apple’s successful implementation of in-house processor designs for select new MacBook and Mac Mini computers could open the door to more ARM-based PCs. Intel’s confirmation that it is outsourcing some production may show some vulnerabilities of the IDM business model. And AI training and inference growing from the datacenter to the edge hints at the next realm of big semiconductor market growth.
The market research and strategy consulting company, Yole invite you to discover today the status of these markets, one year removed from the start of the COVID-19 pandemic.
Traditional CPU terrain seeing more encroachment from ARM-based alternatives
To much fanfare and skepticism, Apple’s rollout of the M1 processor made the processing world take notice. Most reviews of the new ARM-based SoC show it to be a significant achievement, adding doubt to the notion that high-end notebook PCs and even mainstream desktop PCs must contain x86-based CPUs. Publicly, Apple is touting the decision was based on performance enhancements of the M1, though we are confident there is also significant cost savings for Apple as Intel’s CPU margin is no longer a part of the new MacBook and Mac Mini build of materials. Over the next 18 months, Apple has hinted its plans to move their whole PC fleet over to in-house designed processors, discontinuing a lineage of Intel-sourced CPUs dating back to 2006. Yole estimates that Apple’s processor sourcing decision will take ARM-based notebook PCs up from around 10M units in 2019 to more than 30M in 2022.
The next question will be whether other notebook PC OEMs will look to make a similar move. The first obstacle that other OEMs will face is in software; Apple is able to do much more with integration between software and hardware because they have complete control over the Mac ecosystem including the MacOS operating system. For this same reason, iPhones have tended to require less mobile DRAM than their high-end Android counterparts. In a windows-based notebook, hardware/software integration requires significant coordination between separate companies. Even in the case of a Microsoft laptop, where the software decisions are made in-house, any software changes need to accommodate the whole of the Windows ecosystem. The second obstacle that will make Apple’s M1 success harder to replicate is the fact that Apple has been designing their own SoCs for over a decade. While the core IP has been and still is ARM-based, these cores are customized by Apple. Another notebook OEM would have a hard time matching this activity, but they can go to Qualcomm looking for something more powerful than the Snapdragon 8cx, which is already rumored to be under development.
What else is new with Intel?
This January, Intel indicated their intention to shift some CPU production to foundries, with speculation around TSMC taking orders for later in 2021. The implications of this move from Intel can take a few paths and may prove to be a Rorschach test for how you perceive the CPU giant. If one sees Intel as the center of the Logic universe, well on its way towards a goal of $100B of annual revenue ($77.9B in 2020), then utilizing some foundry capacity for lower-end CPUs is a way to optimize the internal manufacturing while making good on commitments to customers. If one is skeptical of Intel’s manufacturing position, prompted by struggles on 10nm (and more recently 7nm), then the move to outsource is further proof that risk sharing between foundries and fabless semiconductor companies is the way of the future. It is this author’s opinion that it is too soon to count out Intel’s process technology prowess, and that any near-term moves to outsource are merely stop-gap measures to keep customers happy while focusing on manufacturing of higher margin products in-house. However, this is one place to keep a close eye, as it will allow an easier contrast between the products designed by Intel and those designed AND manufactured by Intel.
Where is the big growth set to occur in processors?
This brings us to GPUs, which are not just for PC gaming and crypto mining anymore. Recent trends are increasingly placing GPUs into servers. While the GPUs found in servers are not quite like those found in a gaming PC, the suppliers involved and underlying architectures are essentially the same. For a decade, Nvidia may have proven to the industry that the general-purpose GPU (GPGPU), optimized for vector operations could significantly improve AI training and inferencing over a standalone CPU, GPU suppliers are competing with a growing field of AI accelerators to prove which are the best coprocessor for AI model training and inference. The current generation of GPGPUs optimized for AI with tensor cores and other acceleration logic have evolved from just vector to matrix acceleration to compete in this growing field. The difference between client and server markets is stark: server GPUs are riding the wave of AI acceleration, taking on the bulk of the coprocessor workload, but with AI just emerging, other coprocessor designers are looking to find their niche. At least some of the value proposition for choosing Nvidia GPGPU as the server accelerator coprocessor is Nvidia’s massive experience in developing solutions to machine-learning problems, and a wealth of software integration to go with it.
While the portion of servers that use some accelerator coprocessors may never reach above 50%, it is still less than 10% today. As more datacenter usage cases see the benefit of deploying machine learning, then we would expect this portion to grow. At Yole, we have this forecasted to reach 14% of all servers in 2026, of which 80% are satisfied by a GPGPU solution. While AI-specific accelerators are likely to experience the largest growth, their market is still very small in comparison. It will be some time before they can compete with the competitive and well-established, as well as pricey, ecosystem of GPGPUs. This is why we see server GPGPU as the leader in the coprocessor segment, growing revenue three-fold to almost $18B by 2025.
What’s next?
Within the traditional CPU x86 applications, Yole see AMD emerging as a persistent competitor to Intel.
At Yole, we will be watching for further market share gains by AMD, as their predictable architecture and process node cadence gain converts among the customer base. On the datacenter side of things, we will be watching for continued adoption of accelerator coprocessors for AI training and high-performance computing applications.
What will be interesting is how the chips will land between the established GPU players, with their long track-record of software and customer relationships that complement their hardware solutions, and the hundreds of upstarts looking to find a niche in the fast-growing AI ecosystem. Not only will AI be a significant driver in the cloud ecosystem, but it also promises to drive a significant evolution of processor design down to the edge including many APUs and MCUs integrating new AI acceleration architectures.
Completion of the immense acquisition of Arm by Nvidia still looms with significant debate around Nvidia’s ability to sooth government regulators and its pending Arm IP customers. The acquisition could be an AI ecosystem powerhouse for Nvidia and its IP customers, or it could be a boon for the RISC V ecosystem as wary Arm ecosystem customers seek to hedge their bets against the potential of a combined IP supplier and competitor.
Yole’s Processor Quarterly Market Monitor will be published every beginning of March (Q1), June (Q2), September (Q3) and December (Q4). The aim of Yole’s team is to give a closer look at the main markets and players. Analysts invite you to follow our computing activities on i-Micronews, especially during this complex period due to the impact of COVID-19.
Stay tuned to i-Micronews to get further info. about our activities!
John Lorenz
Technology and Market Analyst, Computing and Software
Semiconductor, Memory and Computing Division, Yole Développement
Tom Hackenberg
Principal Analyst, Computing and Software
Semiconductor, Memory and Computing Division, Yole Développement
Adrien Sanchez
Technology and Market Analyst, Computing and Software
Semiconductor, Memory & Computing Division, Yole Développement
Emilie Jolivet
Director
Semiconductor, Memory and Computing Division, Yole Développement
