MiniMax M2.7 AWQ-4bit benchmarks Spark vs RTX

This is a useful reality check on inference economics: brute-force GPU horsepower still wins on speed, but the cheaper Spark setup is closer than the price gap suggests once you factor power and ownership cost.

• –The author reports the 2x RTX 6000 setup at about 2.7x faster on prefill and 4.88x faster on generation, so Spark is not a drop-in replacement if latency is the main KPI.
• –Reported power per 1M tokens is similar enough that the cheaper hardware starts to look attractive for always-on personal or small-team use.
• –The concurrency=2 degradation at high context sizes points to KV-cache pressure and scheduler throttling, not just raw compute limits.
• –This is not a perfectly apples-to-apples bake-off: the Spark cluster is the author’s tuned daily driver, while the RTX box sounds like a quick RunPod/vLLM baseline.
• –For on-prem deployment planning, the lesson is to tune cache format, batching, and serving params before assuming the pricier GPUs will scale linearly.