NVIDIA released the Global Performance Tuner and public implementation in TensorRT 11.1.
NVIDIA opens accuracy-aware tuning for TensorRT build routes
2026-10-24Verified announcementTRIAL
NVIDIA released the TensorRT 11.1 Global Performance Tuner and an open-source trtexec implementation. It searches more than 200 compiler choices across build routes while treating accuracy as a constraint. NVIDIA reports 1.14x to 1.63x gains across ten workloads; independent reproduction is not yet available.
Official · ReferenceTensorRT Global Performance Tuner, Part 1Official · ReferenceTensorRT Global Performance Tuner, Part 2
Availability: open-source trtexec implementation; Python support planned for Torch-TensorRT 2.15
What this means for you
For inference teams, this turns manual builder-route tuning into a repeatable search problem with an accuracy guard. Vendor-only performance numbers and search cost still require workload-specific validation.
Evidence & related material
Technical details & scores
| Version | TensorRT 11.1 |
|---|---|
| Search Space | more than 200 compiler knobs and multiple build routes |
| Metric Boundary | accuracy is a constraint; performance ranks valid configurations |
| Reported Gain | 1.14x to 1.63x across ten publisher-selected workloads |
| Availability | open-source trtexec implementation; Python support planned for Torch-TensorRT 2.15 |
| Time Basis | NVIDIA feed published Part 1 at 2026-10-23T17:27:35Z and Part 2 at 17:27:52Z |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 4
Verification cost 3
Risk 2