**Finding the 4x 3090 Sweet Spot**
A Reddit user conducted a test to understand the efficiency curve of running four RTX 3090 GPUs, each with a power limit. The setup used Qwen3.6-27B (in FP16 mode) and vLLM v0.20.2 backend for benchmarking. The results showed that the optimal power limit is around 220W, where both output throughput and efficiency are maximized at approximately 248 t/s with an efficiency of 1.13 t/joule. This finding aligns with a reference blog post on vLLM performance benchmarks.
**Why It Matters**
This test provides valuable insights into optimizing the power usage for a four-GPU setup, which is crucial for maintaining optimal performance and energy efficiency in AI compute clusters. Understanding these sweet spots can help users like this tester achieve better cost-effectiveness and reliability with their hardware configurations. For those looking to run larger models or more complex workloads on similar setups, this information could guide them towards the best possible power limits.
– The 220W mark is identified as a peak efficiency point.
– Increasing beyond this limit yields diminishing returns in terms of performance improvement but continues to maintain high efficiency.
– Users can now better manage their GPU power settings for optimal performance and cost.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
