Plucky Chinese startup claims to beat Nvidia DGX Spark, Mac Mini on Qwen — says its '256 TOPS beats 1000 TOPS in prefill', running 16-core Cortex-A55 SoC on 40w
Sipeed claims its 256 TOPS mini PC can outperform Nvidia on QwenThe 128GB Sipeed MaixStation costs less than Nvidia's reported $6,949 DGX SparkNvidia advertises 1,000 sparse FP4 TOPS, while Sipeed emphasizes practical inference performance
Chinese electronics manufacturer Sipeed has claimed its upcoming MaixStation AI computer can outperform Nvidia's DGX Spark in certain Qwen workloads despite considerably lower advertised computing performance.
Powered by Axera's AX8860P processor, the compact machine combines 16 Arm Cortex-A55 cores with AI acceleration reaching 256 TOPS and typical consumption of 40W.
Sipeed is also challenging Apple's Mac Mini in its comparisons, although the available benchmark figures primarily concern Nvidia's system rather than Apple's hardware.
MaixStation challenges Nvidia's performance advantage
Sipeed's published Qwen results indicate that its 128GB configuration slightly exceeds the DGX Spark's performance, with further software improvements expected to increase throughput.
The company attributes this outcome partly to differences in how advertised computing capacity translates into actual performance when processing large language models.
Nvidia advertises 1,000 TOPS using sparse FP4 calculations, whereas Sipeed argues that practical workloads cannot fully exploit this theoretical computing capability.
Although Nvidia's dense FP4 performance approaches 500 TOPS theoretically, Sipeed maintains that its architecture uses available processing resources more efficiently during prefill.
Both machines offer 273GB/s memory bandwidth, although Sipeed claims architectural optimizations allow its processor to sustain higher utilization during demanding inference operations.
The comparisons were conducted with speculative decoding through multi-token prediction disabled, allowing the company to evaluate both systems under similar conditions.
However, the reported advantage remains relatively small, and independent testing would be necessary to establish performance differences across additional models and workloads.
The AX8860P uses a 6nm manufacturing process, supports 256-bit LPDDR5X memory and provides 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 containing as much as 128GB memory.
Pricing and hardware configurations
Sipeed expects its entry-level MaixStation with 32GB memory to cost below $2,000, while the 128GB configuration should remain under $4,000.
By comparison, Nvidia's DGX Spark reportedly sells for $6,949.99, considerably above its original $3,000 to $4,000 pricing announced during 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 normal consumption is approximately 40W, the computer receives electricity through a 65W USB Power Delivery connection using its Type-C port.
Sipeed claims its 32GB model supports Qwen3.8-27B, while a 128GB configuration can accommodate the considerably larger Qwen3.8-Flash-Next 125B model.
Combining two units provides 256GB 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 approximately 80% interconnect efficiency when multiple chips operate within connected computing systems.
Software compatibility reportedly includes PyTorch, ONNX, vLLM, SGLang and Hugging Face formats, alongside interfaces compatible with OpenAI applications and Axera's development tools.
The company plans a Kickstarter campaign within several months, offering prospective customers a $100 discount through a reservation costing $30.
Despite its claimed advantages, commercial availability, final pricing, and independently verified performance remain unresolved, particularly for configurations running extremely large language models.
