Embedded World in Nuremberg is an opportunity for processor suppliers and processor- design (IP) licensors to talk about new wares. Hot products this year relate to AI, the RISC-V CPU architecture, and FPGAs.
Key points:
- AI on the rise: New NPUs from ARM (Ethos-U85) and Syntiant (NDP250) cater to edge AI applications.
- RISC-V gains momentum: Imagination Technologies launched the APXM-6200, boasting superior performance density compared to Arm’s offerings. Red Semiconductor’s VISC technology promises to speedup computationally intense functions.
- FPGAs awaken: Altera’s Agilex 5 FPGAs become available, while AMD refreshes its Versal portfolio with enhanced video processing and AI capabilities. Efinix targets traditional FPGA applications with its Ti375 FPGA, which integrates RISC-V cores.
RISC-V
- Imagination Technologies launched the APXM-6200, a multicore-capable RISC-V CPU for application processing (presumably implying a full MMU to run Linux and similar operating systems). Its key attribute is performance density (area efficiency), delivering greater throughput than Arm’s Cortex-A53, Cortex-A55, and Cortex-A510 in less area—the same claim RISC-V rival SiFive makes for its P470 CPU. Better cost is a key RISC-V selling point, making area efficiency an important metric for Imagination and its competitors.
- Red Semiconductor debuted its Versatile Intrinsic Structured Computing (VISC) approach. The concept is as old as processing: special instructions and data paths execute functions faster than general ones. Speeding up math-intensive algorithms like cryptography, VISC mates with RISC-V, a good fit considering the open architecture’s deeply embedded use and customizability.
- X-Silicon disclosed a GPU incorporating a shader core compatible with RISC-V, resulting in a processing unit capable of CPU, GPU, and NPU functions. Think Silicon’s Neox GPU is similar. These designs’ general-purpose processing capability and RISC-V compatibility are only curiosities for now.
FPGAs
- Altera talked up Agilex 5 (midrange) FPGAs, announcing their broad availability and support in the company’s Quartus tools.
- AMD refreshed its Versal portfolio. Like the previous-generation Versals, the new ones integrate much more hard logic than classic FPGAs, incorporating not just video codecs, but also CPU cores, Ethernet and PCIe controllers, a GPU, an image processor (ISP), and other functions. The Prime models upgrade Versal’s video-compression engines, increasing throughput. The AI Edge models support the MX6 block floating-point format, which packs an element’s exponent and mantissa into six bits (compare with FP4 supported by Nvidia Blackwell). These don’t target data centers or high-volume designs but instead industrial, defense, and medical applications where AI adds new functionality and FPGAs implement custom logic.
- Efinix, a young company supplying the classic FPGAs that AMD is deemphasizing, is sampling the Ti375 FPGA, which integrates up to 1 million logic elements, PCIe controllers, and a RISC-V core complete with FPU and MMU. A month ago, the company announced it is targeting the auto industry with qualified parts and tools.
AI
- ARM added a new NPU at the high end of its portfolio, the Ethos-U85, available in configurations from 256 GOPS to 4 TOPS and working with Cortex-M and Cortex-A CPUs. Even 4 TOPS is relatively low throughput, but plenty of embedded applications need this amount or less. Scaling down a higher-performance design may not be effective at reducing power and area; thus, suppliers like Arm must offer purpose-built NPUs for this performance range. Supporting more operators than earlier Ethos models, the U85 should increase real-world performance by more than its OPS speedup and enable it to better handle transformer networks. Arm states the TinyLlama model fully maps to the U85, with no operators falling back to the host CPU. Other changes reduce memory transactions, improve resource utilization, and increase power efficiency. The company also includes the Ethos-U85 in its Corstone-320 design, a preintegrated and preverified combination of Arm IP.
- Syntiant introduced its NDP250 NPU (AI accelerator), integrating the company’s new core 3 engine and quintupling the peak performance of its predecessor. Other cores include an Arm Cortex-M0 CPU and a Cadence Tensilica HiFi 3 DSP. Targeting low-power, always-on designs, the NDP250’s milliwatt scale offers 30 GOPS of number crunching for video and audio applications, such as automotive security and smart doorbells.
Other
- NXP is sampling the S32N55 processor, the first in the S32N series and part of the company’s CoreRide solution, which combines an S32N with independent software for automotive designs. The S32N55 integrates a pile of core Arm Cortex-R52 CPUs, a couple of Cortex-M7 pairs, and security capabilities to enable physical ECU consolidation while maintaining logical separation and isolation. As companies like MediaTek and Qualcomm have entered the auto-semi market, leveraging AI, communications, graphics, and processing investments previously amortized by higher-volume markets (e.g., smartphones), traditional auto semi suppliers must defend their position, especially in ECUs where their processors have dominated.
Looking Forward
We’ll provide more insight as we get additional information on these announcements. The interest that edge (embedded) AI generated several years ago had faded, but we’re starting to see a resurgence as startups shake out and practical applications emerge. The broad edge AI category is served by diverse products, including FPGAs as well as MCUs and MPUs integrating NPUs ranging from tens of GOPS to tens of TOPS. Meanwhile, RISC-V is consolidating disparate architectures used where Arm or x86 compatibility isn’t required. Suppliers with an established customer base and complementary IP have an advantage over startups with limited portfolios but face fierce competition.