Sipeed unveils MaixStation with AX8860P, priced at $3,999 for 128GB model
Sipeed announced the MaixStation, a Chinese-made AI accelerator system built on Axera’s AX8860P chip. The system supports 128GB of LPDDR5X memory over a 256-bit bus at 273 GB/s, priced at around $3,999, matching AMD’s Ryzen AI Halo’s 192GB Max+ PRO 495 model but undercutting Nvidia’s DGX Spark at $6,950. Kickstarter sign-ups are open now, with shipments expected within three months of campaign…
Key points
- MaixStation costs **$3,999** for 128GB LPDDR5X, matching AMD’s **$3,999.99 Ryzen AI Halo 128GB** and undercutting Nvidia’s **$6,950 DGX Spark**
- Axera’s AX8860P chip offers **256 TOPS** at FP4, with **16 Cortex-A55 cores** and PCIe 5.0 support, targeting LLM inference with **25% faster prefill** claims
- Sipeed’s NPU toolchain (AXDSL, Pulsar2) releases by month’s end; Kickstarter launches late 2026 with shipments in early 2027
Sipeed claims the MaixStation delivers 25% faster prefill and 30% faster decode than the DGX Spark when running Qwen 27B models, though Nvidia’s DGX Spark’s dense compute rate (~500 TFLOPS) still exceeds the AX8860P’s 256 TOPS at native FP4. The AX8860P, a 16-core Arm Cortex-A55 with PCIe 5.0 support, is produced in a 6nm process. Sipeed also highlights its 256 Gbps C2C interconnect (scalable to 384 Gbps) for multi-node setups, contrasting with Nvidia’s QSFP-based Spark, which requires external switches for scaling. The MaixStation ships with Debian 13 and Linux 6.18, with Windows support pending demand, and supports formats like Hugging Face, ONNX, and an OpenAI-compatible API.
The story so far
6 episodes →- Sipeed unveils MaixStation with AX8860P, priced at $3,999 for 128GB modelthis story
Sipeed reveals Chinese-made DGX Spark and Ryzen AI Halo rival, expected to cost $3,999
Tom's Hardware · 15 October 2026
Loading the full article…
This text was published by Tom's Hardware and written by Shane Downing. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Mindbeam and NVIDIA launch CUDA kernel generator and inference engine for Nemotron 3.5 Super VL · 3 src
- Nvidia speeds DLSS 5 with 2x performance, 20% less VRAM · 4 src
- Dogness launches AI Pet Feeder with Google Gemini for behavioral monitoring · 1 src
- Portland-based Mueon, which says its AI server racks could fit up to 10x more computing capacity, raised $65M led by Maverick Silicon, with Intel participating · 1 src
- Amazon launches Bee Frontier wearable and upcoming Bee Eclipse with subscription · 4 src
Comments
via GitHub Discussions