VRAMora

VRAMora provides a visual comparison platform for local large language model (LLM) inference hardware, allowing users to analyze and filter options based on system cost, memory capacity, speed, and power consumption. It supports a range of hardware types, including NVIDIA and AMD GPUs, Apple Silicon, packaged systems, Mac Mini, and clusters. The tool is designed to help users identify suitable hardware for running LLMs locally by presenting data in an interactive chart where the X-axis represents system or GPU-only cost, and the Y-axis can be switched between metrics such as maximum model size, VRAM, token speed, or efficiency (tokens per second per watt).

Users can interact with the chart to view bubble sizes indicating token generation speed at a specific model and quantization (7B Q4), and bubble colors reflecting power efficiency, with green for efficient systems and red for power-hungry ones. Hovering over a bubble reveals detailed specifications, including cost, memory, speed, power draw, maximum model size, and additional notes. Sidebar filters enable narrowing results by hardware type and memory tier, while a GPU-only cost mode is available for those comparing the incremental cost of upgrading a GPU in an existing system. The platform also features a Y-Axis dropdown to switch between key performance and efficiency metrics, and a highlight function to visually single out the best value, fastest, most efficient, or highest memory options among filtered hardware.

A model selection dropdown allows users to pick an LLM and instantly see which hardware configurations have sufficient memory to run it, with visual cues for compatibility. For detailed analysis, VRAMora offers a sortable table view of all hardware data. The platform includes a sharing feature that encodes current filters and view settings into a URL, enabling users to share specific comparisons. Benchmark data is sourced from community testing and public resources, with figures such as system costs and token speeds provided as approximate estimates. The platform is web-based and includes mobile-responsive design for usability across devices.

VRAMora serves those evaluating or purchasing hardware for local LLM inference, providing a detailed, interactive environment for comparing performance, cost, memory, and energy efficiency across a range of hardware options.