# NVIDIA Details TensorRT Tuner’s 1.14x to 1.63x Performance Across 10 Workloads

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/39838
Published: 2026-10-24T09:52:49.000Z
Updated: 2026-10-24T09:52:49.000Z
Section: Technology

> The Global Performance Tuner option was added to `trtexec` in TensorRT 11.1, released June 24, 2026.

NVIDIA detailed Friday, Oct. 23, how its TensorRT Global Performance Tuner searches alternative engine build routes and selects the fastest candidate that stays within an application’s accuracy limit. Across 10 workloads, the fastest qualifying route delivered between 1.14x and 1.63x the default route’s end-to-end performance. Each workload’s default route was normalized to 1.0x, making the figures relative measurements rather than absolute latency or throughput.

A build route combines values for compiler knobs, settings that influence how TensorRT builds neural-network inference engines for NVIDIA GPUs. TensorRT 11.1 offers more than 200 tunable compiler knobs. With no route specified, TensorRT uses default values. A route does not define every part of the engine: TensorRT also makes build decisions, including tactic selection, and may adjust requested values to satisfy network constraints.

In trtexec, the tuner expands route expressions into candidate builds, measures them and removes candidates that crash or fail accuracy checks. Developers set an acceptable accuracy boundary for their application, so the filter reflects that application’s requirement. Candidates that pass are ranked by end-to-end performance, and the highest-performing eligible build is selected. Speed determines the ranking only after failed or inaccurate candidates have been excluded.

The tuner offers fast, mixed and full search strategies. Fast varies one knob at a time after a default-route baseline; mixed follows that scan with combinations of promising settings. Full evaluates every combination in the search, and the number of combinations grows exponentially as more knobs vary. The breadth of a full search therefore increases quickly when developers add more settings to the search.

The Global Performance Tuner option was added to trtexec in TensorRT 11.1, released June 24, 2026.

Results depend on the model, input shapes, target GPU, software stack, and TensorRT and tuner versions. A build route that improves one deployment can have no effect or lower performance in another. The 1.14x-to-1.63x range reflects the tested workloads, not a single expected gain across all systems.