Axera AX8860P 16-core Arm SoC targets desktop AI at under $4,000

Axera unveiled the AX8860P, a 6nm 16-core Cortex-A55 SoC with up to 256 TOPS (FP4) AI performance and 273 GB/s memory bandwidth. Sipeed will use it in the MaixStation, a 110×110×30 mm USB-C/10GbE node with 32GB or 128GB RAM, priced under $2,000 and $4,000 respectively. Radxa is also developing a Pocket Supercomputer with the chip. Sipeed claims the 128GB MaixStation matches or slightly exceeds NVIDIA's DGX Spark on Qwen benchmarks while costing significantly less, and plans a Kickstarter launch in coming months.

AX8860P specifications and Sipeed hardware

• CPU: 16x Arm Cortex-A55 cores
• AI accelerator: Up to 256 TOPS (FP4 dense), 128 TOPS otherwise
• Memory: Up to 128GB 256-bit LPDDR5x at 273 GB/s bandwidth
• Video: 64-channel decode, 16-channel encode
• I/O: PCIe 5.0, AXLink (C2C) chip-to-chip cascading at 80% efficiency
• Power: ~40W under load in 78×78 mm module
• Software: vLLM, SGLang, PyTorch, ONNX, Hugging Face, OpenAI-compatible APIs; Axera toolchain with axmodel conversion from ONNX
• MaixStation: 32GB or 128GB LPDDR5x, USB 3.0 Type-C host, 10GbE SFP+, 65W USB PD, 110×110×30 mm
• Scaling: Single 32GB node runs Qwen3.8-27B; 128GB node runs Qwen3.8-Flash-Next 125B; dual 256GB node handles DeepSeek V4-Flash 284B; quad 512GB node manages DeepSeek V4.1-Flash 552B

Why it matters

If Sipeed's pricing and performance claims hold, the MaixStation could significantly lower the cost of entry for local LLM inference and small-scale AI clusters, particularly for developers already using PyTorch and Hugging Face workflows. The use of standard USB-C and 10GbE for daisy-chaining avoids expensive InfiniBand switches.