2x3090 hits 250W sweet spot for Qwen3.6-27B

Local inference enthusiasts facing the 3090 "power wall" should prioritize 250W for efficiency, though 275W offers a measurable latency win for single-stream requests.

• –Speculative decoding via Multi-Token Prediction (MTP) is the primary performance driver, delivering a massive 2x throughput boost for coding tasks.
• –Tensor parallelism (TP=2) is mandatory to mitigate PCIe bandwidth bottlenecks on dual-GPU consumer setups, outperforming single-card configurations.
• –The combination of FP8 KV caching and INT4 AutoRound quantization enables a massive 200K context window within the 48GB VRAM limit.
• –Benchmark results confirm that LLM inference remains memory-bandwidth bound; aggressive 350W stock power limits generate excess heat with negligible token gains.