This week, the annual Consumer Electronics Show offered a big stage for processor suppliers to present new products. An emphasis on AI capabilities continues in everything from game GPUs to microcontrollers. Below are observations about major processor-related disclosures. Highlights include AMD’s Strix Halo CPU-GPU combo and Nvidia’s GB10-based workstation.
AMD
The most significant AMD announcement was the official Strix Halo (Ryzen AI Max) unveiling. Construction resembles other AMD PC processors, integrating two computing dice (CCDs) and an I/O die (IOD) but stuffing a big GPU into the IOD. Unlike AMD’s other chiplet-based processors, Halo appears to employ a silicon substrate instead of connecting the dice with an ordinary organic interposer, which should improve interdie throughput. Its powerful graphics obviate a discrete GPU in designs requiring more performance than past integrated-graphics processors (IGPs or APUs). AMD, thus, is likely to take share from Intel (CPU) and Nvidia (GPU) while reducing system cost.
Conspicuously absent was an RDNA 4 (Radeon RX 9000) announcement, which AMD at the last minute deferred to later in the month. Considering how important graphics are to the company’s PC and custom game-console designs, it’s unserious about the discrete GPU business. AMD has long trailed Nvidia in introducing new features and in recent years has failed to capitalize on its rival moving so far upmarket as to be unaffordable for many consumers. Nonetheless, the opportunity for AMD to play Volkswagen to Nvidia’s Mercedes remains available.
AMD Gets Krackan
AMD also disclosed its first Krackan Point PC processors, sold alongside Strix Point as Ryzen AI processors. Integrating fewer Zen 5 and GPU cores, Krackan addresses lower-cost designs. Other revealed PC processors were specific models of existing dice or close cousins thereof. For example, the chiplet-based Fire Range (Ryzen 9000HX) for laptops is similar to the desktop Granite Ridge (Ryzen 9000X), and the Ryzen Z2 series for handheld games comprises variations of various laptop processors. The Ryzen 200 processors are Zen 4 AMD Hawk Point Refresh chips, which are as discernable an update as Intel Raptor Lake Refresh was.
Ordinarily we don’t discuss design wins, but AMD finally landed sockets in Dell’s commercial (business) PCs. Despite succeeding in gaming systems, AMD’s PC-processor share has long lagged that of Intel. This gap will close with AMD gaining Dell’s business and garnering wins at other major PC OEMs.
Nvidia
Whereas AMD deferred revealing its new GPU, Nvidia debuted Blackwell PC GPUs (RTX 50). Sharing a name and architectural attributes with the data-center Nvidia Blackwell GPU, these deliver new features and greater performance and set a new high-water mark for GPU list prices: $2,000. The company’s glib performance comparisons with its prior-generation GPUs were quickly debunked and overshadowed the new features. Nvidia has led the industry in applying AI techniques to graphics, employing them initially to generate interstitial frames and upscale images. Shader code can now invoke small neural networks, adding AI capabilities to the graphics-processing pipeline. Nvidia GPUs, for example, can apply AI to improve ray-traced scenes and render materials and faces better than before. Such features may take a while for developers to adopt, and Nvidia is seeding the market to enable them to begin.
Arm and Digits
Nvidia’s other big hardware disclosure was the small Project Digits workstation based on the new GB10 “superchip,” a module combining scaled-down versions of the Grace Arm processor and Blackwell data-center accelerator connected by the company’s NVLink interface. Integrating an Nvidia ConnectX network adapter and supporting the NCCL protocol, Digits systems can pair up to run bigger models.
As with the RTX 50, Nvidia glibly characterized the system’s performance, sometimes being unclear about employing FP4 to achieve a particular throughput. Devs will likely use Digits for creating and training models and will favor FP16 and greater-precision types. Nonetheless, at $3,000, the system is inexpensive considering its capabilities and AI-developer target market. When it ships in May, it should be cheaper than an x86 workstation with an add-in card and faster (and closer to deployment targets) than the matrix extensions built into recent Xeons. An XPU truism is that the processor with the most developers ultimately prevails, and Digits will help Nvidia maintain its lead.
Nvidia disclosed few GB10 details but stated MediaTek helped develop it, likely designing the 20-core Grace processor. The two companies have other collaborations, including an ongoing Windows-on-Arm design that may share technology with the new Grace. MediaTek also likely supplies the Digits wireless and USB connectivity chip.
Other Nvidia News
Nvidia disclosed numerous nonhardware AI developments related to transportation, industrial automation, smart retail, and other areas. Although large language models are AI mania’s biggest fuel source, AI’s potential goes much further. Nvidia has the most to gain from AI proliferation and the wherewithal to invest liberally in many industries.
Intel
Intel announced several PC-processor series, but all employ the previously disclosed Lunar Lake and Arrow Lake chips. The company’s processors dominate the Windows PC market, and Intel finally has a complete product stack based on current-generation process technology. The company is sampling new processors fabricated in its 18A process technology and expects volume production in 2H25. It didn’t clarify if this timeline applies to Panther Lake laptop processors, Nova Lake (desktops), or Clearwater Forest (servers), but it did demo Panther systems.
Previous disclosures indicate that the Intel 18A process includes gate-all-around (GAA) transistors like the stillborn 20A technology and backside power. Intel could be the first to mass produce wafers with the latter, which offers several advantages, including greater logic density. Successful ramping of 18A production would signal the end of Intel’s manufacturing woes, enable the company to declare 5N4Y a triumph, and help Intel Foundry Service (IFS) land big customers. The organization has already stated Amazon will build a chip using Intel 18A. A fabric adapter (not a switch fabric, like Intel suggested), the design has low stakes for Amazon but an important win for IFS.
Regarding products, Panther and Nova are likely chiplet-based designs employing a silicon substrate like their predecessors Meteor Lake and Arrow Lake. Reports stating 18A will account for 70% of Panther/Nova production indicate that Intel is likely to continue employing TSMC to make its GPU chiplets. Relying on TSMC for chiplets would suggest IFS can’t meet even the internal GPU team’s requirements, auguring ill for its development as a foundry for external customers. Note also that the year before Arrow launched, it appeared as an Intel 20A design instead of the TSMC-fabricated product that Intel ultimately ramped. Thus, while Intel promises to make Panther and Nova internally, a lot can change in the coming months. It’s essential to the company (really, the whole industry) for 18A to succeed, and the chipmaker’s promise to dogfood it is a good sign. But it’s only that, a sign.
Embedded Suppliers
Outside the computing sphere, CES had other processor-related news.
- Alif disclosed new AI-accelerated MCUs. Sitting at the high-performance end of the microcontroller spectrum, they integrate an Arm Ethos-U85 engine and Cortex-M55 CPUs with various multimedia functions to enable AI processing of images, sound, and other sensors. A startup chasing the MCU market, Alif faces stiff competition from established suppliers, such as STMicroelectronics, which in December released the M55-based STM32N6 with AI acceleration.
- Ambarella announced it’s sampling the N1-655 vision and AI processor. It’s a lower-cost and -power version of previous N1 chips but still employs high-performance, safety-enhanced Arm Cortex A-78AE CPUs. The killer AI application in industrial and other embedded settings has been object recognition and video processing. A longtime leader in classical computer vision chips, Ambarella has added AI acceleration to the N1 line, including the ability to handle transformer models.
- Renesas and Honda signed an agreement to codevelop a 3 nm SoC for automotive electrical control units. Employing chiplets, the design combines Renesas’s R-Car X5 with a Honda AI accelerator. This effort shows the modularity benefits that chiplet designs offer and how established auto semi companies are advancing even as new entrants like Qualcomm and Nvidia seek to leapfrog them.
- Google endorsed the MediaTek MT7903 wireless combo chip for smart home devices. It’s not a processor but a complement. MediaTek’s cost advantages (lower opex and lower manufacturing costs owing to tight physical designs), fast development pace, and advanced wireless technologies position it above western companies such as NXP, TI, and Microchip in the smart home. Offering Wi-Fi 6E and BT6 along with Thread, the MT7903 will be a technological step ahead of competitors’ combo chips.
- NXP announced its intent to acquire TTTech Auto. Part of the Austrian TTTech company family, TTTech Auto focuses on software and system engineering for cars. Because defining a processor like those NXP sells requires understanding how it will be used and “enablement” (software, services, and support) to get customers to select it, the small company is a valuable acquisition for NXP.
- Qualcomm made numerous announcements about partnerships and customer success in smart home, automotive, and industrial electronics. New hardware includes an enterprise system based on the Cloud AI 100 accelerator, which has changed little from 2020. The company also disclosed new Qualcomm Snapdragon X SKUs for inexpensive PCs.