A user detailed their experience benchmarking a two-machine setup for running Llama models via RPC, focusing on the performance differences across various network links. The user encountered significant issues with Thunderbolt 5 connectivity, finding it unreliable for DMA-based transfers and ultimately opting for a USB 10Gbps connection. Benchmarks revealed that Thunderbolt 5 performed worse than a standard 5GbE connection in terms of latency and throughput, while USB 10Gbps offered slightly better performance, though with some model-specific exceptions. The testing involved six different models and explored both single-stream and concurrent request handling, noting that memory overflow became an issue for denser models at higher concurrency levels. AI
IMPACT Highlights potential hardware bottlenecks for efficient local LLM deployment.
RANK_REASON User benchmark of hardware connectivity for AI model serving.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →