爱芯元速 AX8860P:16 核 A55 SoC 以零头价格挑战英伟达 DGX SparkAxera AX8860P 16-core Cortex-A55 SoC takes on NVIDIA DGX Spark at fraction of the cost


AI free token/AI免费积分
爱芯元速 AX8860P:16 核 A55 SoC 以零头价格挑战英伟达 DGX Spark

【产品开发】爱芯元速(Axera)创始人兼 CEO 邱晓欣博士近日在中国国际工业博览会上发布了全新的 AX8860/AX8860P 处理器。这是一款 16 核 Cortex-A55 SoC,NPU 算力最高 128/256 TOPS(FP4),配备 256 位 LPDDR5/5X 内存与 PCIe 5.0 互联,目标是让桌面级 AI 方案以英伟达 DGX Spark 零头的价格与之竞争。CNX Software 对这款芯片及配套硬件进行了详细报道。

AX8860P 主要规格:CPU 为 16 核 Arm Cortex-A55;AI 算力最高 256 TOPS;VPU 支持 64 路解码、16 路编码;内存最高 128GB 256 位 LPDDR5x,带宽 273GB/s;互联采用 PCIe 5.0 + AXLink(C2C)级联,多芯片卡与系统级联效率 80%;负载功耗约 40W;78×78mm 模组封装;6nm 工艺。

软件生态方面,该 SoC 支持 vLLM、SGLang、PyTorch、ONNX、Hugging Face 格式、OpenAI 兼容 API 以及爱芯自家的工具链。官方文档已演示如何把 MobileNetV2 ONNX 模型转换为运行在爱芯芯片上的"axmodel"。

硬件伙伴方面,Radxa 将推出基于 AX8860P 的"口袋超算",但细节尚未公布;Sipeed 则带来多款 AX8860P PCIe 卡(MaixBlade 最高四芯片)与 MaixStation 迷你 AI 电脑——后者现已开放预订。MaixStation 规格:32GB 或 128GB LPDDR5x 内存,USB 3.0 Type-C 接笔记本或 10GbE SFP+ 接局域网,65W USB PD 供电,典型功耗 40W,尺寸 110×110×30mm。

算力实测方面:单台 32GB 节点可运行 Qwen3.8-27B;单台 128GB 节点支持 Qwen3.8-Flash-Next 125B;双节点 256GB 可跑 DeepSeek V4-Flash 284B;四节点(MaixStation Pro)512GB 可管理 DeepSeek V4.1-Flash 552B。Sipeed 公布的 Qwen 基准显示 MaixStation 128GB 仅略快于 DGX Spark 128GB,但公司预计后续优化会扩大差距。Sipeed 解释称,"256 TOPS"之所以能打赢英伟达宣传的"1000 TOPS",是因为后者是稀疏 FP4 算力、真实模型用不上,Dense 纸面约 500 而 Spark 只发挥了一小部分,MaixStation 则能把 256 TOPS 更充分地利用起来。

价格是最锋利的卖点:DGX Spark 2025 年 3 月首发时 3000-4000 美元,如今已涨到 6949.99 美元;MaixStation 尚未开售,Sipeed 称 32GB 版将低于 2000 美元、128GB 版 3999 美元封顶。多机级联成本差距更大——MaixStation 只需一根线缆,而英伟达方案需要 4000 美元的 200G/400G 交换机。Sipeed 将在几个月内登陆 Kickstarter,用户支付 30 美元预订费即可在众筹启动时抵扣 100 美元。

本文来源:https://www.cnx-software.com/2026/10/19/axera-ax8860p-16-core-cortex-a55-soc-takes-on-nvidia-dgx-spark-at-fraction-of-the-cost/
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Axera AX8860P is a new 16-core Cortex-A55 SoC with up to 128 TOPS/256 TOPS (FP4) NPU, 256-bit LPDDR5/5X memory, and PCIe 5.0 connectivity, aiming to power desktop AI solutions competing with NVIDIA DGX Spark at a fraction of the cost.

Dr. Qiu Xiaoxin, founder and CEO of Axera, unveiled the new AX8860/AX8860P processor a few days ago at the China International Industrial Fair, along with upcoming products from Sipeed and Radxa.

Axera AX8860P key features and specifications: CPU - 16x Arm Cortex-A55 processor; AI - Up to 256 TOPS; VPU - 64-channel decode, 16-channel encode; Memory - Up to 128GB 256-bit LPDDR5x with 273GB/s bandwidth; Interconnect - PCIe 5.0 + AXLink (C2C) cascading with 80% efficiency for multi-chip cards and systems; Power Consumption - About 40 Watts under load; Package - 78 x 78 mm module; Process - 6nm.

The AX8860P SoC is said to support vLLM, SGLang, PyTorch, ONNX, Hugging Face formats, OpenAI-compatible APIs, and Axera's toolchain. Some documentation is already available showing how to convert a MobileNetV2 ONNX model into an "axmodel" running on the Axera chip.

In terms of hardware, Radxa will introduce a "Pocket Supercomputer" powered by the AX8860P SoC, but has yet to provide details since it's still a few months away.

Axera AX8860P 16-core Cortex-A55 SoC takes on NVIDIA DGX Spark at fraction of the cost

Sipeed also has a few AX8860P PCIe cards on the way (with up to four chips with the MaixBlade), and a MaixStation mini AI computer for which we have more details, and you can even reserve one. MaixStation specifications: SoC - Axera AX8860P; System Memory - 32GB or 128GB LPDDR5x; Host interface - USB 3.0 Type-C port connected to a laptop, 10GbE SFP+ cage connected to the local network; Power Supply - 65W USB PD via USB Type-C port; Power Consumption - 40W typical; Dimensions - 110 x 110 x 30 mm.

A single 32GB node can run Qwen3.8-27B, a single 128GB node supports Qwen3.8-Flash-Next 125B, a dual node with 256 GB RAM can handle DeepSeek V4-Flash 284B, and a quad-node (Maixstation Pro) with 512GB manages DeepSeek V4.1-Flash 552B.

Sipeed also shared Qwen benchmarks comparing the MaixStation 128GB to the DGX Spark 128GB. The numbers show it's only slightly faster, but they expect further optimization will widen the gap. The company explains that "256 TOPS" beats the "1,000 TOPS" advertised by NVIDIA because the latter is sparse FP4, which real models can't use, while Dense is around 500 on paper, but Spark only puts a small slice of it to work, while MaixStation makes better use of its 256 TOPS.

The bandwidth is the same for both systems at 273 GB/s, but Axera AX8860P is said to implement architecture improvements to "keep more of those 273 GB/s busy". Finally, MTP (speculative decoding) is turned off in the benchmarks above for side-by-side comparison.

The NVIDIA DGX Spark was first unveiled in March 2025 for $3,000 to $4,000, but since then, price increases have kicked in, and it sells now for $6,949.99. While the MaixStation is not for sale now, the company says the 32GB model will go for under $2000 and the 128GB model for under $4000, let's make that $3,999. Daisy-chaining more than two systems is also much cheaper on the MaixStation since you only need a cable, but a $4000 200G/400G switch for the NVIDIA systems.

Sipeed will launch it on Kickstarter in a few months and invites users to pay a $30 reservation fee to get 100 off once the crowdfunding campaign starts.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry