Specifications
Products
Axelera Embedded 110m and Axelera Embedded 113m
AI Engine
Quad-core Metis AIPU (AI Processing Unit)
Peak Throughput
Up to 214 TOPS per card
Architecture
Digital In-Memory Computing (D-IMC) with RISC-V controlled dataflow
Host Interface
PCIe Gen 3.0 x4
Form Factor
NGFF M.2 add-in card, 22 x 80 mm class
Dedicated Memory — 110m
1 GB DRAM dedicated to the AIPU
Dedicated Memory — 113m
Up to 8 GB dedicated memory on the 113m
Typical Power
5 to 9 W typical application power
Target Workloads — 110m
Efficient, performance-enhanced vision inference
Target Workloads — 113m
LLM and VLM workloads plus multi-camera vision inference with cascaded or parallel models
Security
Secure Boot and Root of Trust integrated on-card
Thermal Options
Extended temperature options available; passive and active cooling variants
Toolchain
Voyager SDK with quarterly software updates improving delivered performance
Model Support
TensorFlow, PyTorch and ONNX models
Host Compatibility
Adds AI acceleration to existing x86 or Arm systems without redesigning the enclosure
Design Intent
Right-size inference hardware — no GPU power or thermal budget required
Related PCIe Card
Axelera also offers a single-slot HHHL PCIe CXP-class card with up to 4 GB or 16 GB memory at 8 to 15 W
Deployment Speed
Up to 3,200 FPS ResNet-50 on the PCIe variant
Supply
Global B2B supply via QS Compute — quote on request
Overview
Axelera's Embedded product line is a family of industry-standard M.2 AI inference cards that exist so that edge systems can right-size their acceleration instead of over-provisioning a GPU. Both the 110m and the 113m are powered by the same quad-core Metis AIPU, and both deliver up to 214 TOPS while drawing only about 5 to 9 W in typical application use.
The two cards differ by intent. The Embedded 110m is the efficient vision-inference part: one Metis AIPU with 1 GB of dedicated DRAM, minimal power, and the simplest possible integration into an NGFF M.2 socket to boost an existing vision pipeline. The Embedded 113m adds capacity and workload range — up to 8 GB of dedicated memory, support for LLM and VLM workloads, and multi-camera inference running cascaded or parallel models, with secure boot and extended temperature options.
The underlying Metis architecture combines proprietary Digital In-Memory Computing with RISC-V controlled dataflow and is programmed through the Voyager SDK, which Axelera updates quarterly with performance improvements that carry to hardware already deployed. For applications that need a slot rather than a socket, Axelera also builds a single-slot HHHL PCIe card on the same silicon, with 4 GB or 16 GB of memory, 8 to 15 W typical power, active air cooling, and published throughput up to 3,200 FPS on ResNet-50.
QS Compute supplies the Axelera Embedded card family for evaluation and for production, along with the companion Voyager SDK entitlement path and carrier integration support.
Key Benefits
214 TOPS in a 5 to 9 W envelope: inference density without a GPU thermal budget. M.2 drop-in: upgrades an existing x86 or Arm vision system without a chassis redesign. Dedicated memory on card: 1 GB on the 110m and up to 8 GB on the 113m keeps inference traffic off the host DRAM. Secure Boot and Root of Trust: hardware-anchored trust for regulated edge deployments.
Applications
Multi-camera machine vision and inspection systems, intelligent video analytics and people counting, retail and smart-city edge appliances, LLM and VLM assistants embedded in industrial equipment, robotics perception, medical imaging instruments, and retrofits that must add AI inference to an already-deployed x86 or Arm platform.
Request a Quote — AXELERA EMBEDDED 110M AND 113M M.2 AI ACCELERATOR CARDS — METIS AIPU AT 5 TO 9 W
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →