Axera's AX8860P claims to outpace Nvidia DGX Spark in select LLM benchmarks
Axera introduced the AX8860 series, with the AX8860P chip featuring 16 Arm Cortex‑A55 cores, a 6 nm NPU delivering up to 256 TOPS of FP4 compute, 128 GB LPDDR5X memory support, 273 GB/s bandwidth, PCIe 5.0 and AXLink interconnects. The chip uses Axera’s proprietary AXNeutron data‑flow architecture, which the company says improves utilisation for transformer‑based language models.
Key points
- Axera's AX8860P provides 256 TOPS FP4 compute, 16 Cortex‑A55 cores, 6 nm process, 273 GB/s bandwidth.
- Sipeed says its AX8860P‑based MaizStation can outperform Nvidia DGX Spark on selected LLM inference tests.
- MaizStation will launch via Kickstarter end‑2026, priced under $2,000 for 32 GB and under $4,000 for 128 GB models.
Sipeed, a partner building the MaizStation compact AI computer around the AX8860P, claims its device can beat Nvidia’s DGX Spark on certain LLM inference benchmarks, citing native dense FP4 acceleration and better memory bandwidth use. The benchmarks are supplied by the manufacturer and have not been independently verified. The MaizStation will be offered with 32 GB or 128 GB of LPDDR5X memory, priced below $2,000 and $4,000 respectively, with a Kickstarter launch planned for late 2026. Radxa is also developing an AX8860P‑powered Pocket Supercomputer.
The story so far
7 episodes →- Axera's AX8860P claims to outpace Nvidia DGX Spark in select LLM benchmarksthis story
Axera AX8860P takes aim at Nvidia DGX Spark with proprietary AI accelerator
notebookcheck.net · 19 October 2026
Loading the full article…
This text was published by notebookcheck.net and written by Anil Ganti. 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 →- Opinion: Google’s AI focus leaves Pixel missing Samsung’s Good Lock and widget stacking · 1 src
- Nvidia projected to sell about $430 billion in chips in 2026 as data-center revenue hits $89 billion · 6 src
- Microsoft could make Maia 300 its primary inference engine with 2.5x memory bandwidth · 2 src
- Nvidia’s GB300 FP4 submission trails Google’s TPUv7 on cost efficiency · 1 src
- Tenstorrent raises $1.5B Series E at $6.2B valuation led by Trace Capital and Innovation X · 3 src
Comments
via GitHub Discussions