Geekbench AI is a benchmarking suite with a testing methodology for machine studying, deep studying, and AI-centric workloads, all with the identical cross-platform utility and real-world workload reflection that our benchmarks are well-known for. Software program builders can use it to make sure a constant expertise for his or her apps throughout platforms, {hardware} engineers can use it to measure architectural enhancements, and everybody can use it to measure and troubleshoot gadget efficiency with a set of duties based mostly on how units truly use AI.

Geekbench AI measures your CPU, GPU, and NPU to find out whether or not your gadget is prepared for at this time’s and tomorrow’s cutting-edge machine studying purposes. It contains each pc imaginative and prescient and language checks that mannequin real-world machine studying duties and purposes.

What’s New

Geekbench AI 1.0 contains different vital adjustments to reinforce its means to measure real-world efficiency based mostly on how purposes are utilizing AI. This contains assist for brand new frameworks, from OpenVINO on Linux and Home windows to vendor-specific TensorFlow Lite delegates like Samsung ENN, ArmNN, and Qualcomm QNN on Android to higher mirror the newest instruments out there to engineers and the altering ways in which builders construct their apps and companies on the most recent {hardware}.

This launch additionally makes use of extra in depth knowledge units that extra intently mirror real-world inputs in AI use instances, and these greater and extra various knowledge units additionally enhance the effectiveness of our new accuracy evaluations. All workloads in Geekbench 1.0 run for no less than a full second, which adjustments the affect of vendor and manufacturer-specific efficiency tuning on scores, making certain units can hit their most efficiency ranges throughout testing whereas nonetheless reflecting the bursty nature of real-world use instances.

Extra importantly, this additionally higher considers the delta in efficiency we see in actual life; a five-year-old telephone goes to be a LOT slower at AI workloads than, say, a 450W devoted AI accelerator. Some units may be so extremely quick at some duties {that a} too-short take a look at counter-intuitively places them at a drawback, underreporting their precise efficiency in lots of real-world workloads!