Primate Labs is pleased to announce the availability of Geekbench AI 1.1, the latest update in our cross-platform AI inference workloads benchmark.
This release includes:
- Framework and runtime upgrades - This includes an upgrade to ONNX Runtime 1.19 on Windows, improving Half Precision support on AMD and Intel CPUs, and changes to the Core ML configuration, which improves performance on iOS 18 and macOS 15. OpenVINO now falls back to the CPU in Single Precision workloads if the device does not support the required data types rather than convert to a different data type at runtime. ArmNN has been upgraded to v24.08, and Samsung’s ENN to v3.1.8.
- Better results on some Android systems - Some Android devices suffered a bug through the Play Store deployment that limited the hardware Geekbench AI had access to. Extracting bundled libraries to the file system mitigates the issue, improving performance.
- Score validation improvements - By randomly re-validating extra iterations in a small percentage of cases, the robustness of the benchmark is improved. Validation is now parallelized in most workloads, reducing the time spent verifying results — this means the benchmark should take less time to run.
- Requantize all the things - LOTS of models for LOTS of supported frameworks have been requantized, improving output quality, accuracy, and performance across workloads. Expect to see score increases.
- Smaller tweaks - Geekbench AI now uses per-image normalization ranges in Depth Estimation, improving accuracy calculations, and our intern adjusted some dates in the source code — give them a big hand, please.
On the balance, most device scores will be slightly higher with this release, meaning Geekbench AI 1.1 scores aren’t strictly compatible with Geekbench AI 1.0 scores. For any test feedback, questions, or concerns, reach out to us on Discord or via email.
Geekbench AI 1.1 is a recommended update for all Geekbench AI users.