Lucebox has partnered with AMD to demonstrate a heterogeneous consumer hardware setup that outperforms the Nvidia DGX Spark in inference speed. The system combines an AMD Radeon AI PRO R9700 GPU with a Strix Halo processor to run the full 284B DeepSeek V4 Flash model.
- The configuration achieves a decode speed of 51.1 tok/s on the complete model.
- It delivers a 3.63x performance advantage over a single Nvidia DGX Spark unit.
- The R9700 handles the dense path, hot experts, cache, and draft model, while Strix Halo manages other experts in 128GB of memory for simultaneous computation.
- The setup costs $2,899 less than purchasing two Nvidia DGX Spark devices.
The authors note that this is an experimental implementation running a Q2 model at 16k context, with plans to implement KVFlash to support 64k-128k contexts.