AMD Strix Halo runs AI for $0.48/day, sips power compared to Nvidia A6000
A Reddit user reveals Strix Halo's energy cost beats Nvidia for local AI inference.
In a viral Reddit post, user Forward_Jackfruit813 highlights the extreme energy efficiency of AMD's Strix Halo APU for local AI workloads. They report a worst-case daily power cost of just $0.48 when running multiple models, compiling software, and hitting CPU, GPU, and NPU simultaneously for 24 hours. Under silent, power-sipping conditions, the device handles 50 tokens per second on a Qwen 3.6 35B model in Q8_XL quantization—all in a form factor as small as a router.
The comparison to an Nvidia A6000 is stark: the A6000 consumes 300W for the card alone, more than double the entire Strix Halo system's power budget. The user emphasizes that beyond inference, any unused compute can host services, making it a versatile all-in-one package. While acknowledging Nvidia cards are faster, they argue that for professionals concerned about size, noise, energy use, or running multiple tasks, Strix Halo punches well above its weight—even at current inflated prices.
- Strix Halo draws at most $0.48/day in worst-case 24/7 workload (compiling + multi-model inference).
- Silent operation delivers 50 tokens/second on Qwen 3.6 35B Q8_XL, matching a small router's footprint.
- A single Nvidia A6000 GPU consumes 300W, more than double the Strix Halo's entire system power budget.
Why It Matters
Professionals running local AI can drastically cut electricity costs and space while retaining useful inference speed.