TensorRT Global Performance Tuner
A TensorRT 11.1 feature that builds candidate engines over multiple build routes and picks the fastest one meeting an accuracy criterion.
Overview
NVIDIA introduced the Global Performance Tuner in TensorRT 11.1, described in a two-part series dated October 23, 2026. Instead of one fixed scheduling strategy, it generates candidate engines through customizable build routes, checks each candidate's outputs against a user-specified accuracy criterion, measures end-to-end inference performance and selects the fastest candidate that passes. NVIDIA frames deployment as a compilation problem with many valid solutions. Part 1 covers the principles (accuracy-aware tuning, customizable routes, large solution space) and Part 2 the algorithm, architecture and best practices. The posts report no benchmark figures.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Multiple customizable build routes per model
- Accuracy-aware selection against a user-defined criterion
- End-to-end latency measurement of each candidate engine
- Shipped in TensorRT 11.1; documented on docs.nvidia.com
Use cases
- Squeezing inference latency out of a model without sacrificing accuracy thresholds
- Automating TensorRT engine tuning per deployment environment
Pricing
unknown; part of TensorRT, pricing not stated.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 23 Oct 2026
- Record updated
- 29 Oct 2026
- Record last checked
- 29 Oct 2026