Try the wait before buying a GPU with llmspeed's local AI simulator

A more tangible way to compare local AI hardware

llmspeed turns local-model speed estimates into a streaming preview you can watch before committing to a GPU or Mac. An October 17 Show HN submission brought fresh discussion to the existing browser tool, which lets visitors select hardware, a model and quantization, then compare two configurations in a race.

The useful distinction is between a throughput number and the experience of waiting. The application includes an estimated delay before the first token as well as the subsequent text stream. It also checks memory fit, including configurations that require partial offload or cannot run. The README lists 30 hardware choices and 15 models; those are the project’s catalog counts, not a claim of comprehensive hardware coverage.

This is a simulator, not a local inference test. Choosing a configuration does not run that model on the visitor’s computer, and the streamed preview does not establish its answer quality.

Read the label before trusting the race

The project separates Measured, Derived and Projected estimates. Its README defines the first as a benchmark for the exact hardware, model and quantization, the second as interpolation from nearby measured setups, and the third as a calculation from specifications, including announced but unreleased hardware.

When a direct benchmark is absent, the documented approach uses memory bandwidth for generation, compute for prompt processing, and backend coefficients calibrated against community measurements. Memory estimates include weights, cache and runtime overhead. The live interface exposes a calculation breakdown rather than presenting every configuration as an equally well-established measurement.

Those distinctions matter for shoppers. In the English HN thread, one commenter reported that their M2 Ultra performed much faster than the tool’s estimate for a different M3 Ultra configuration; another asked for clearer inference-library and version information. These are user reports and requests, not controlled comparisons or proof of a specific prediction error. AITrending has inspected the application and documentation but has not reproduced its hardware benchmarks.

The README limits the simulator to a single device and warns that drivers, engine versions, thermals and other settings affect actual results. It is a useful starting point for comparing the feel of candidate setups, followed by checking measurements for the exact hardware and runtime being considered. The new hook is today’s public discussion, not a claim that this older project launched today or has demonstrated rapid growth.