Radxa Unveils the AXERA AX8850-Powered AICore AX-M1 M.2 Edge AI Accelerator
Gumstick-format device packs eight Arm Cortex-A55 cores, a 24 TOPS NPU, vision processor, and 8GB of dedicated RAM.
Embedded and hobbyist hardware specialist Radxa has announced an upcoming M.2 M-key accelerator for on-device machine learning and edge artificial intelligence workloads, built around the AXERA Tech AX8850 and delivering a claimed 24 tera-operations per second (TOPS) of INT8 precision compute: the Radxa AICore AX-M1.
"The Radxa AICore AX-M1 is a high-performance M.2 computing acceleration module designed for edge AI computing and inference applications," Radxa explains of the gumstick-format board. "Powered by [an] AXERA [Tech] AX8850 SoC [System-on-Chip], [with] integrated octa-core [Arm] Cortex-A55 CPU. Built in acceleration for transformer-based model inference, up to 24 TOPS at INT8 [precision]. Uses [a] standard M.2 2280 M-key form factor, compatible with mainstream industrial and embedded motherboards."
The M.2-format accelerator, brought to our attention by Linux Gizmos, is designed to offer a drop-in performance boost for on-device machine learning and artificial intelligence tasks - and uses AXERA Tech's latest AX8850 chip, which packs eight Arm Cortex-A55 running at 1.5GHz alongside a vision processing unit offering hardware encoding and decoding of H.265/H.265 video at up to 8k30 or 16 channels at 1080p30, an in-house neural processing unit with a claimed compute performance of 24 TOPS at INT8 precision, and 8GB of dedicated LPDDR4X memory.
The company says that the accelerator can be used on host devices with AMD, Intel, and Rockchip processors — with no word yet on whether it's compatible with a Raspberry Pi 5 using the M.2 HAT+ — and supports acceleration of a range of popular models including DeepSeek, Qwen, Gemma, Llama, Phi, YOLO, Whisper, Stable Diffusion, Real-ESRGAN, and more. On the software front Radxa says it's compatible with Canonical Ubuntu Linux and upstream Debian Linux, though it shouldn't be difficult to extend that support to other distributions.
More information on the AICore AX-M1 is available on the official product page; at the time of writing, Radxa had not confirmed pricing nor availability.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.