Science & Tech · AI · 5 hrs ago
Sipeed says its MaixStation can beat Nvidia’s DGX Spark on Qwen
Chinese electronics company Sipeed says its upcoming MaixStation can outperform Nvidia’s DGX Spark on some tasks running Qwen language models.
The compact computer uses an Axera processor with 16 Arm cores, AI acceleration rated at 256 TOPS and typical power use of 40 watts.
Sipeed’s published results show a slight lead on a Qwen workload, although the company says the comparison focuses on practical performance rather than Nvidia’s higher advertised TOPS figure.
Both systems have the same stated memory bandwidth, and the tests disabled a feature called speculative decoding.
Sipeed expects the MaixStation to cost under $2,000 for 32GB of memory and under $4,000 for 128GB, compared with a reported $6,949.99 price for Nvidia’s system.
Independent tests are still needed to confirm the claimed advantage across other models and workloads.
Sipeed plans to launch a Kickstarter campaign within several months, but final pricing and availability have not been confirmed.
Chinese electronics manufacturer Sipeed says its upcoming MaixStation can outperform Nvidia’s DGX Spark on some Qwen workloads.
The MaixStation uses Axera’s AX8860P processor, with 16 Arm Cortex-A55 cores and AI acceleration rated at 256 TOPS.
Sipeed’s published results show its 128GB configuration slightly ahead of the DGX Spark, but the advantage has not been independently verified.
The company says the MaixStation’s lower advertised computing capacity can deliver strong practical performance during large language model inference.
Sipeed expects the 32GB model to cost under $2,000 and the 128GB model under $4,000.
- Who
- Sipeed, a Chinese electronics manufacturer, and Nvidia.
- What
- Sipeed says its upcoming MaixStation can outperform Nvidia’s DGX Spark on some Qwen workloads.
- When
- The MaixStation is upcoming; Sipeed plans a Kickstarter campaign within several months.
- Where
- Not stated
- Why
- Sipeed says its processor uses available resources more efficiently during inference, especially in prefill.
Sipeed
Nvidia
Benchmark performance
Sipeed
Sipeed says its 128GB MaixStation slightly exceeds the DGX Spark on published Qwen results.
Nvidia
Nvidia advertises 1,000 sparse FP4 TOPS for the DGX Spark; the article gives no Nvidia response to Sipeed’s comparison.
Computing capacity
Sipeed
Sipeed argues that practical workloads cannot fully use Nvidia’s advertised sparse FP4 capacity.
Nvidia
Nvidia advertises 1,000 sparse FP4 TOPS, while its dense FP4 performance is described as approaching 500 TOPS theoretically.
Nvidia announced original DGX Spark pricing of $3,000 to $4,000.
Sipeed plans to launch a Kickstarter campaign for the MaixStation.
- Processor
- Axera AX8860P
- Cores
- 16 Arm Cortex-A55 cores
- AI acceleration
- 256 TOPS
- Typical power consumption
- 40W
- Expected price, 32GB model
- Under $2,000
- Expected price, 128GB model
- Under $4,000
What people are saying
256 TOPS beats 1000 TOPS in prefill