A home lab enthusiast successfully ran the 2.8-trillion-parameter Kimi K3 AI model using Intel Optane Persistent Memory 200 Series modules. The setup relies on CPU-only inference, as the model typically requires data center hardware and at least 610GB of RAM to function.
The builder utilized second-generation Optane sticks in App Direct Mode to manage the massive memory capacity at a lower cost than standard DRAM. Intel launched these modules in 2020 for 3rd-gen Xeon Scalable platforms, offering up to 32% more bandwidth than the first generation.
Performance reached 0.75 tokens per second with 3-bit quantization. The system generates significant heat, with each module consuming 12W to 15W. To prevent damage to adjacent DRAM, the builder implemented blower fans to maintain the required airflow of 2.4 to 2.7 meters per second.
Thermal throttling for these modules begins at 83 degrees Celsius, with a hard shutdown at 85 degrees. The creator intends to optimize the system by adjusting thread counts and incorporating GPUs and DRAM as a faster caching tier for future benchmarks.
