AMD revealed its embedded-processing strategy at its recent analyst day and set growth targets. Less noticed than data-center and client businesses, the embedded unit is called out in the company’s quarterly financial statements. It faces the same growth pressures but unique dynamics. The diverse embedded market is in disarray at the high end, and AMD’s strategy falls short of capitalizing on the market’s opportunities.
Defining Terms
Strictly defined, an embedded system is a microprocessor-based design for which end users may not load software. Put another way, it’s any microprocessor-based design that isn’t a computer. The processor used in an embedded system is an embedded processor. Such definitions are academic, however. System categories are imprecise, and chip categories are even fuzzier. The same chip counts as a computer processor when used in a server and as an embedded processor when used in a storage appliance.
AMD muddies the water further by reporting most of its FPGA sales in the embedded unit. Moreover, its Zynq and Versal products fuse an FPGA and a complete processing subsystem. Putting them in one category or the other is an arbitrary choice. Sales of Ryzen and Epyc chips for embedded systems account for most of the remainder. The Epyc line includes 8004-series Siena models, which have smaller packages than other Epycs and are more power efficient. They target edge servers and similar applications that can be categorized as embedded.
AMD’s Analyst Day
In its analyst-day presentation, AMD broke its embedded business into four pieces:
- Adaptive computing (FPGAs)
- Embedded x86 (Ryzen and Epyc)
- Semicustom products
- Physical AI
The first two are the company’s familiar products. In its presentation, AMD asserted it has expanded its FPGA line to cover more low-cost, low-density devices. However, its long-extant Spartan product line was overdue for a refresh to integrate up-to-date memory and I/O interfaces. For x86 processors, the company had applied distinct model numbers for PC processors retargeted at embedded designs (e.g., Ryzen V3000). It now simply uses the same model numbers, only adding Embedded to the name (e.g., Ryzen Embedded 9950X). Thus, the company has expanded its embedded lineup without developing new silicon, whereas an embedded-specific design would better serve the market
The semicustom business is new and builds on its experience with the PlayStation and Xbox. Semicustom is a vague term that could apply to special bins of high-volume chips, packaging changes, or designs with custom logic or chiplets. The company reports having secured $15 billion worth of semicustom designs since 2022, divided roughly evenly among wireless, aerospace and defense, data center, and automotive customers.
A term Nvidia popularized, physical AI refers to machines operating in the real world (meatspace) and employing AI. For example, a robot arm (a type of embedded system) can employ AI to classify objects within its field of view and direct its operation. The physical AI market will grow because of a combination of de novo growth (e.g., from humanoids) and the reclassification of cars, industrial robots, and other familiar systems.
AMD is on track in 2025 to land embedded designs worth $16 billion, indicating wins are now growing at a 21% annual rate compared with the previous 8% rate. The company expects to grow its 55% FPGA market share to 70% in 2030, a remarkable expansion. Over the next 10 years, physical AI will expand the addressable market.
State of the High-End Embedded and FPGA Markets
For the past few years, embedded revenue has been in a cyclical downturn for the whole industry. For high-end processors and FPGAs specifically, the product market has been in disarray for longer.
As we previously described, Intel chased away embedded-processor competitors targeting communications applications, the area that had driven performance and innovation. Intel’s faltering execution and competitors’ exits have left this market underserved. The company’s plan to spin out its Network and Edge Group affects the business’s long-term viability because it would separate the business from its source of CPU designs.
AMD, however, doesn’t offer standard alternatives to Intel Xeon D or Atom P5000, two products targeting communications designs. A semicustom design could replace the former, but the P5000 (Snow Ridge) requires more tech (IP) than AMD has in-house.
Intel also amassed customers preferring to base their designs on PC-like hardware instead of on purpose-built embedded processors. The company’s success here has further dissuaded erstwhile competitors and has become a multibillion-dollar annuity. Fortunately for AMD, this success has also opened the market to an x86-compatible alternative. However, it remains Intel’s to lose.
Dominated by AMD (Xilinx) and Intel/Altera, high-density FPGA development slowed when Intel acquired Altera. The two companies’ FPGAs were among the first semiconductors to employ the latest process technology. Both companies have introduced new products in the past few years, but neither matches the previous new-product cadence or employs a leading-edge process.
The embedded-processor market, however, is still reeling from the Zynq introduction, which enlarged Xilinx’s addressable market and propelled revenue. Since then, Xilinx (later, AMD) has made Arm-based processing subsystems a common feature. The recent Versal Prime Gen 2 devices, for example, integrate eight Cortex-A78AE cores. Such integrated devices obviate two-chip FPGA-plus-microprocessor designs. However, this direction was set long ago, before AMD’s acquisition of Xilinx. Accelerating growth requires doing something new.
AMD Promises More of the Same
Beyond supporting semicustom designs, AMD’s embedded strategy doesn’t break fresh ground but extends its historical approach. Without an innovation such as Zynq, the company’s FPGA business is in zero-sum competition with Intel. Versal’s non-FPGA blocks or similar IP could be employed in a standalone multicore processor, for example.
The embedded x86 business, likewise, is opportunistic, not strategic, only retargeting PC and server chips. The company could create fresh silicon better aligned with embedded requirements but still compatible with PC/server system software. Even if it chooses to only keep using PC/server chips, AMD must develop more partners and be aggressive to capture market share.
In contrast to AMD’s embrace of the status quo, other companies are taking fresh approaches to the embedded market. Qualcomm, for example, is entering the automotive market. The company initially targeted car cockpits with adaptations of its smartphone chips and now has a portfolio addressing an array of automotive functions. Nvidia is pursuing the broad embedded market, offering Arm-based chips with AI-processing capabilities, which it supports with modules and an AI-focused software stack.
Bottom Line
The embedded market has many opportunities, and AMD has an enviable technology portfolio. The unsettled competitive landscape and the potential for infusing AI throughout the embedded market present an unusual circumstance for AMD to become a dominant force. Seizing the opportunity, however, requires more than refreshing old product strategies and pursuing a few semicustom designs.