RAGMark: per-stage RAG benchmarking shows reranking plus compression cuts energy 66%

_reachsumit · x · 2026-09-09

RAGMark is a modular benchmarking framework for RAG systems in small multi-GPU setups, profiling retrievers, vector databases, prompt processing, and generators with per-stage latency, GPU utilization, memory, power, TTFT, throughput, and answer quality metrics. Across five QA workloads it shows autoregressive generation dominates latency in naive pipelines, while context-reduction shifts bottlenecks across stages; reranking and compression compound — jointly cutting prefill and KV-cache traversal costs and lowering energy use by up to 66%.

Original post →

More from Infra

Infra channel →