Specifications
Brand
Axera
SoC
Axera AX8850
CPU
8× Arm Cortex-A55, up to 1.7 GHz
NPU
24 TOPS INT8
Memory
8 GB 64-bit LPDDR4x @ 4266 Mbps (4 GB system + 4 GB acceleration on some SKUs)
Storage
32 GB eMMC 5.1 (boxed units); 32 Mbit QSPI NOR boot flash on the M.2 card
Video Encode
H.264 / H.265 up to 8Kp30
Video Decode
H.264 / H.265 up to 8Kp60, up to 16× concurrent 1080p streams
Host Interface
PCIe 2.0 ×2 on the M.2 card; M.2 M-Key 2242
Boxed Variants
M5Stack AI Pyramid (4 GB) and AI Pyramid-Pro (8 GB, adds HDMI input)
Module Variant
M5Stack LLM-8850 M.2 2242 with turbine fan and CNC aluminium heatsink
Card Variant
Radxa AICore AX-M1 M.2 accelerator
Networking (boxed)
2× Gigabit Ethernet
Display (boxed)
HDMI 2.0 output up to 4Kp60; HDMI 2.0 input on the Pro
USB (boxed)
4× USB-A 3.0, 2× USB Type-C (one with PD 3.0 up to 100 W)
Dimensions (boxed)
approx. 114.5 × 105.0 × 62.0 mm, 195 g
Framework Support
Linux-first; ONNX and standard inference runtimes
Overview
The Axera AX8850 is an edge AI SoC that pairs an octa-core Arm Cortex-A55 cluster at 1.7 GHz with a 24 TOPS INT8 NPU and a video pipeline capable of 8Kp30 encode, 8Kp60 decode and sixteen concurrent 1080p decode streams. That combination targets two jobs at once: local model inference and multi-channel video processing on the same silicon.
It reaches the market in three shapes. The M5Stack AI Pyramid and AI Pyramid-Pro are small boxed computers — the Pro adds HDMI input and doubles memory to 8 GB — with dual Gigabit Ethernet, USB 3.0 and a distinctive pyramid enclosure. The M5Stack LLM-8850 is an M.2 2242 card with 8 GB of LPDDR4X and PCIe 2.0 ×2, aimed at Raspberry Pi 5, Rockchip RK3588 SBCs and x86 mini PCs with a spare M-key slot. The Radxa AICore AX-M1 is the same SoC in a comparable M.2 accelerator format.
Compared with vision-only NPUs, the AX8850 is positioned for generative and multimodal work as well as detection: language models, vision-language tasks and real-time object detection all run locally, with video transcoding offloaded to the same device. For NVR and analytics boxes the sixteen-stream decode is often the deciding spec.
Deployments are Linux-first. The M.2 card draws from a host over PCIe and keeps its own 8 GB of memory for model runtime, so the host's RAM budget is unaffected; the boxed AI Pyramid units are standalone and need only power and network.
Key Benefits
24 TOPS INT8 with an 8K video pipeline covers inference and multi-channel transcoding on one device · M.2 2242 card form factor upgrades Raspberry Pi 5, RK3588 and x86 hosts without a platform change · dedicated onboard memory keeps model runtime off the host · boxed and module variants let the same silicon serve both kiosk and retrofit deployments.
Applications
Multi-camera video analytics and NVR transcoding, on-premise edge LLM and vision-language inference, retail and industrial vision inspection, smart-building people counting, and AI acceleration retrofits for existing ARM SBC or mini-PC fleets.
Request a Quote — AXERA AX8850 — 24 TOPS EDGE AI PLATFORM (AI PYRAMID, LLM-8850, AICORE AX-M1)
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →