Qwen3.8-27B on one RTX 3090 vs two: +20% decode, +14% cold prefill

Tests show Qwen3.8-27B at W4A16 fits on a single RTX 3090 and decodes at 125–155 tokens/sec. Splitting it across two cards with tensor parallelism adds about 20% to decode and 14% to cold prefill, while cached prompts answer 2–3x faster.

• One RTX 3090: 123–154 tokens/sec; two cards: 146–193 tokens/sec
• Cold prefill: 1,100–1,220 tokens/sec on one card vs 1,230–1,360 on two
• TTFT on an 8,600-token prompt: 8.5 s on one card, 6.9 s on two
• The 2–3x speedup on repeated prompts comes from prefix caching, not the cards