johnfink8/sfnx (sFNX) is a CUDA inference engine that divides each layer of a model's computation across multiple PCIe graphics cards. According to the repository description, it works on three PCIe V100s, splitting the layers so that a single request reads each card's memory concurrently. It implements the addition between cards directly, using the PCIe connection without a dedicated card-to-card interconnect module. The README states that it produces 382 tokens per second (tok/s) on this setup.
The project provides a separate engine and serving layer for running MoE LLMs across multiple cards. It also includes OpenAI- and Anthropic-compatible interfaces. The repository contains an image processor along with design, benchmark, and validation documents; the implementation language is Python and the license is MIT. The README publishes model execution and concurrent conversation handling results for a specific V100 configuration. So the point of interest is not single-card performance but the way layers are split across multiple PCIe-connected cards.
That said, 382 tokens per second is the result for the specific three-V100 configuration the repository describes. The README explains the measurement method along with the limits of its applicability, and notes that a CUDA setup, model files, and run configuration are required. There are also constraints on the supported graphics card generations and precisions. No evidence is provided that the same results appear with other card configurations or models, so the performance figure should be read within that hardware and measurement scope.