At a glance
| Booth | R8000 |
| Country | NL |
| Website | axelera.ai |
Company profile
Axelera AI brings AI inference out of the cloud and closer to where data is generated and used, covering applications from advanced vision to large language models and from generative AI to robotics. Its Embedded, Edge, and Server product lines run on AIPUs that Axelera describes as the world's most efficient, supporting deployment across a range of environments. Axelera AIPUs are purpose-built processors specifically designed for AI inference. The company targets machine-learning system developers and integrators, software vendors, and OEMs with AI inference acceleration hardware in industry-standard form factors and a supporting software stack. Axelera positions its hardware around performance per watt and performance per dollar, with the goal of simplifying development, integration, and deployment of AI inference solutions. Axelera AI names Ultralytics, Advantech, and SECO among its ecosystem partners; these collaborations include YOLO models running on Metis and Metis integration into Edge AI or other product lines.
Exhibits
Axelera has announced three AIPU versions for different use cases, form factors, and performance requirements: Metis, Europa, and Titania. Metis delivers 214 TOPS within a 4-8 watt budget and is offered in compact M.2 modules or PCIe cards for real-time edge inference. Metis supports up to 16GB of memory per chip and up to 8B-parameter language models per chip, while its architecture is optimized for vision with flexibility for local language models. Europa delivers 629 TOPS, incorporates on-die pre- and post-processing, doubles the AIPU core count relative to the prior design, adds 16 vector processing units, and scales from heavy-edge systems to enterprise data centers. Europa supports up to 64GB of memory per chip, up to 32B-parameter language models per chip, and up to 100B parameters across multiple chips. Titania is described as a chiplet that scales Axelera's Digital In-Memory compute architecture toward high-performance computing and AI Gigafactories. Across the hardware family, listed card and system form factors include M.2, PCIe cards, compute boards, chiplets, PCIe boards, OCP-like insert cards, NUCs, PCs, workstations, edge servers, enterprise servers, rack-mount servers, and supercomputers. Axelera Embedded 110m uses one Metis AIPU with 1 GB of dedicated DRAM and is intended for AI acceleration in constrained systems with an NGFF M.2 socket. Axelera Embedded 113m uses Metis with up to 8 GB of dedicated memory and is presented for LLM, VLM, and multi-camera vision inference on edge and embedded devices. Axelera Edge 130p is a half-height, half-length PCIe accelerator powered by one Metis AIPU, delivering up to 214 TOPS for demanding vision applications and local language models. The Edge 130p is specified at up to 3,200 FPS for ResNet-50, and one published example lists 865 FPS on Ultralytics YOLOv5s versus 543 and 334 FPS for two competitor accelerators. Axelera Edge 232p is a half-height, half-length PCIe card based on Europa, delivering up to 629 TOPS and 83.4 tokens per second for QWEN 30B. The Edge 232p is designed for agentic AI and generative AI from Physical AI use cases through enterprise language serving, with support for 100B-parameter models. Axelera Server 150p uses four quad-core Metis AIPUs and is specified for peak performance of 856 TOPS, up to 12800 FPS for ResNet-50, and up to 38884 FPS for MobileNet V2-1.0. Axelera Embedded 111c combines a quad-core Metis AIPU with an ARM-based RK3588 CPU in a single-board computer and provides PCIe, USB, HDMI 2.0, dual Gigabit LAN, GPIO, UART, I2C, 64GB eMMC, SATA, and microSD connectivity or storage options. Voyager SDK supports both Metis and Europa, using YAML configuration to compile, optimize, and deploy complete pipelines for vision models, image processing, and language models from one software environment. The Edge 130p and Edge 232p use Voyager SDK for evaluation, software integration, and advanced quantization, with end-to-end AI pipelines and models available out of the box. Axelera Edge cards are designed as plug-and-play upgrades for existing infrastructure, with passive-cooling and active-cooling configurations available. A single Metis PCIe card is demonstrated processing 24 concurrent live camera streams with YoloV5 object detection for multi-stream video analytics. Another demonstrated use case performs real-time monitoring of personal protective equipment such as helmets and vests for workplace safety. The Edge product line consists of industry-standard PCIe accelerator cards that combine software ease of use, power efficiency, and scalability for inference acceleration. A single Edge board can run inference from multiple cameras while supporting multiple independent neural networks in parallel. Metis combines Digital In-Memory Computing and RISC-V controlled dataflow technology with Voyager SDK and includes built-in data security. Europa includes 8 second-generation AI cores, 16 RISC-V vector processing cores, 128MB of L2 SRAM, 200GB/s of DRAM bandwidth, onboard pre- and post-processing, and a video decoder to reduce CPU load. A separate offline demonstration shows Axelera Edge 130p running a 3B-parameter Llama 3.2B-Instruct model with zero cloud dependency and 100% local processing. For Ultralytics YOLOv8s, the AIPU comparison table lists 653 frames per second for Metis and 2850 frames per second for Europa, with the latter shown as 4.36x. Axelera Embedded 111c is paired with the Voyager toolchain for model deployment and optimization and is described as supporting PyTorch, ONNX, and TensorFlow workflows. Europa is described as delivering up to 6x more tokens per second per watt than legacy GPU solutions for Physical and Enterprise AI workloads. Voyager Wingman is identified as Axelera's agentic platform for conversational creation of pipelines and deployment troubleshooting.
Capabilities and products
- AIPU AI inference processors
- Embedded, Edge, and Server AI inference product lines
- PCIe edge AI inference accelerator cards
- Voyager SDK AI application compilation, optimization, and deployment