- Sipeed claims its 256 TOPS Mini PC can beat Nvidia on Qwen
- The 128GB Sipeed MaixStation costs less than Nvidia’s reported $6,949 DGX Spark
- Nvidia announces 1,000 sparse FP4 TOPS, while Sipeed emphasizes practical inference performance
Chinese electronics maker Sipeed has claimed that its upcoming MaixStation AI computer can outperform Nvidia’s DGX Spark in certain Qwen workloads despite significantly lower advertised computing performance.
Powered by Axera’s AX8860P processor, the compact machine combines 16 Arm Cortex-A55 cores with AI acceleration that reaches 256 TOPS and typical consumption of 40W.
Sipeed also challenges Apple’s Mac Mini in its comparisons, although the available benchmark figures mainly concern Nvidia’s system rather than Apple’s hardware.
MaixStation claims Nvidia’s performance advantage
The Sipeed published Qwen results show that its 128GB configuration easily exceeds the performance of the DGX Spark, with further software improvements expected to increase throughput.
The company attributes this result in part to differences in how advertised computing capacity translates into actual performance when processing large language models.
Nvidia advertises 1,000 TOPS with sparse FP4 calculations, while Sipeed argues that practical workloads cannot fully exploit this theoretical computing capacity.
Although Nvidia’s dense FP4 performance theoretically approaches 500 TOPS, Sipeed maintains that its architecture uses available processing resources more efficiently during prefill.
Both machines offer 273GB/s memory bandwidth, although Sipeed claims that architectural optimizations allow their processor to maintain higher utilization during demanding inference operations.
The comparisons were carried out with speculative decoding disabled by multi-token prediction, which allows the company to evaluate both systems under similar conditions.
However, the reported benefit remains relatively small, and independent testing would be needed to establish performance differences across additional models and workloads.
The AX8860P uses a 6nm manufacturing process, supports 256-bit LPDDR5X memory and offers PCIe 5.0 connectivity alongside Axera’s proprietary AXLink interconnect.
Its specifications also include 64 video decoding channels, 16 encoding channels and support for configurations with as much as 128GB of memory.
Prices and hardware configurations
Sipeed expects its entry-level MaixStation with 32GB of memory to cost under $2,000, while the 128GB configuration will remain under $4,000.
In comparison, Nvidia’s DGX Spark is reportedly selling for $6,949.99, significantly above its original $3,000 to $4,000 prices announced in March 2025.
The compact AI Mini PC measures 110 × 110 × 30 mm and includes USB 3.0 Type-C connectivity alongside a 10GbE SFP+ network interface.
Although the normal consumption is around 40W, the computer receives power through a 65W USB Power Delivery connection with its Type-C port.
Sipeed claims its 32GB model supports Qwen3.8-27B, while a 128GB configuration can accommodate the significantly larger Qwen3.8-Flash-Next 125B model.
The combination of two units offers 256GB of memory for DeepSeek V4-Flash 284B, while a four-unit MaixStation Pro configuration offers 512GB for DeepSeek V4.1-Flash 552B.
AXLink supports communication between processors, with Axera claiming around 80% interconnect efficiency when multiple chips operate within connected computing systems.
Software compatibility includes reporting PyTorch, ONNX, vLLM, SGLang and Hugging Face formats, in addition to interfaces compatible with OpenAI applications and Axera development tools.
The company is planning a Kickstarter campaign within a few months, offering potential customers a discount of $100 through a reservation that costs $30.
Despite its claimed advantages, commercial availability, final prices, and independently verified performance remain unresolved, especially for configurations running extremely large language models.
About CNX software