Axelera AI

Where is edge compute heading as multi-modal vision and agentic models move into production? At AI Infra Summit 2026 (Santa Clara, CA | Sept 15–17), the conversation centers on solving real-world latency, energy constraints, and operational control in physical deployment. Key presentation to add to your agenda: "The Physical AI Inflection Point" Fabrizio Del Maffeo, Co-Founder & CEO addresses how purpose-built hardware architectures are bridging the gap between cloud-trained models and mission-critical edge performance. What you can evaluate live at Booth #930: - Multi-Tier Edge Inspection: See how pairing baseline continuous tracking with targeted high-tier inspection optimizes throughput and thermal budgets. - Real-Time Video Analytics: Experience natural language video query execution running directly on hardware without cloud dependency. - Agentic Edge Workflows: Test how Voyager Wingman streamlines model compilation, optimization, and edge deployment for fast developer turnaround. If you’re architecting enterprise servers, autonomous robotics, or real-time computer vision systems, explore performance benchmarks and power efficiency with our team in Santa Clara. Reserve a 1-on-1 session or hardware demonstration: https://lnkd.in/ehRPzT4v #AIInfraSummit #PhysicalAI #EdgeAI #Semiconductors #ComputerVision #SystemArchitecture

Sanket Shah recently set about deploying ZipDepth, the zero-shot monocular depth model from CVPR 2026, on his Axelera hardware. One camera with no stereo rig, and no lidar, but a dense depth value for every pixel at 30 FPS on 1080p video. Impressive approach, too. Our compiler wouldn't swallow the whole graph, so he split ZipDepth into four stages, two on the AIPU and two on the host, and chained them so the output of one compiled graph feeds straight into the next. The AIPU stages cover 95% of the model and run at 110 FPS. The CPU head is what holds the end-to-end pipeline at 30, and he's already mapped out how to move it onto the accelerator and clear 60! He also fed a non-image embedding tensor directly into a compiled model and got a result numerically indistinguishable from FP32, which isn't something we see attempted often. Best of all, the write-up is honest about the challenges. The region fusion constraint that kept breaking the graph, a DMA aliasing segfault, OpenMP threads fighting the AIPU for memory bandwidth, and the calibration run that clipped the depth range until he swapped random data for 300 real frames from the actual scene. It's the sort of work that makes obstacle avoidance, floor plane estimation and bin picking possible on a single sensor. Repo's Apache 2.0, and there's more robotics perception on the way. Full write-up, and the discussion around it right here: https://lnkd.in/eqMhSWdP

Axelera's Alexis Crowell was on NYSE Wired recently, and offered a great argument for edge AI compressed into one sentence. It's easy to miss, but well worth putting on a post-it note if you're building with AI right now. "A legal team building contracts doesn't need a model that was trained on French poetry." For a while we've held the assumption that capability scales with parameter count, so the answer to any problem is more compute. 😄 But most production workloads are actually quite narrow. Whether it's a contract review, defect detection on a manufacturing line, or a vision system watching a supermarket floor. Match the model to the job and the economics change, because you stop paying for all that capability and compute you never even use. This is held true in the real deployments we're seeing, too. The blocker is almost never raw model capability. It's cost per inference, latency, and where the data is legally allowed to go. Solve those, and suddenly the deployment looks less and less like a rented data center. Full session is well worth a watch, and thanks to Brian J. Baumann for sharing. 👍

Last shout on The Prompt, Axelera's latest AI hackathon. We're reading the applications right now, but if you're super fast while we're doing it, you can still pitch a project and be one of ten builders getting a Dell Technologies Pro Slim Plus XE5 with a 16GB Axelera PCIe card, free Voyager Wingman, and a month to build your project. €1,000 for the winner. Go be a maverick! https://lnkd.in/eWHiiQ-U

Everyone knows about the first man on the moon. But how often do we talk about the spacecraft launched back in 1977 that is still traveling through interstellar space - farther than anything else humanity has ever built? When it came to naming our company, software, and chip architectures, we wanted that exact same spirit of endless ambition. Here’s the story behind our names, straight from our CEO and Co-Founder, Fabrizio Del Maffeo. Read the blog with the full video in the comments. #AIAccelerators #AIHardware #AISoftware #DeepTech #SpaceExploration #TechInnovation

WG Tech Solutions Pvt Ltd cut worker safety violations by 28% in a working factory using 45 AI models working in unison, not just one. So we asked the WG Tech team, who are regulars in our community, for the engineering decisions behind it. Why they deliberately didn't standardise on a single card. How the models are spread across the Axelera hardware range. What they'd do differently next time. It's valuable, inspiring stuff, so take a look at how it was done: https://lnkd.in/eaP2HnhB

Ten people are about to have a Dell Pro Slim Plus XE5 with a 16GB Axelera PCIe card delivered to their door. Free. Theirs to keep. All they did was pitch an idea for an awesome AI project they want to build. That's The Prompt. You tell us what you'd build using Voyager Wingman, our AI pipeline generator, and we pick the ten most inspiring project ideas. Selected builders get the Dell Technologies XE5, free Wingman access, and one month to build. €1,000 for the winning project, €500 for the runner-up. Wingman installs in a few minutes straight from a web chat, so you're building on real Axelera silicon the same day the box arrives. The YAML and the wiring are its job. The awesome ideas are yours. Pitches are closing soon, so don't hang around: https://lnkd.in/ebyRGy9M

Same silicon. Same performance. Clearer names. Starting today, Axelera AI products carry a new naming system built around three descriptors — Embedded, Edge, and Server — plus a number that tells you the silicon generation, optimization target, relative performance tier, and form factor at a glance. Nothing about the architecture, the roadmap, or what's shipping today has changed. Metis® M.2 is now Axelera® Embedded 110m. Metis PCIe (4-chip) is now Axelera Server 150p. Same boards, same specs, names that finally match how customers actually evaluate hardware. And "optimized for" is a starting point. If an Edge card is the right fit for a server chassis, that's your call. Read the full breakdown: https://lnkd.in/eKbh2hvv #EdgeAI #AIHardware #Semiconductors #AIatEdge #ComputerVision #AIAccelerator

We're #hiring a new VP - People in Zurich. Apply today or share this post with your network.

The World Cup might be over, but one of our engineers spent an evening building his own goal-line technology, on the floor of his living room, out of an Axelera stress ball and two plants in mugs. Roll the ball at the "goal" and the system calls GOAL or MISS. Rebounds off the wall don't count, because the ball has to actually cross the line. Three layers, no custom training, just a stock Ultralytics yolo26s finds the ball as a "sports ball", a nearest-point tracker follows its center and bridges the gaps when detection blinks out on a fast, blurry shot, then a referee layer checks whether the path crossed the line between the posts in the scoring direction. The good bit is what happened to the bottleneck. On his laptop, YOLO detection ate 94% of the processing time. On Metis, detection dropped to 37% and his own OpenCV drawing and video encoding became the slow part, at 63%. Detection and decoding hit around 250 fps on the card, with the whole annotated pipeline running at 94. Full write-up, GIFs and all: https://lnkd.in/eShPyRzc