Benchmark scores for the Google Tensor chip have started to appear online as early adopters and media have received their Pixel 6 and 6 Pro devices. Although Google highlights the ML capabilities of the Tensor chip, synthetic benchmarks such as GeekBench ML do not show promising results. The Apple A15 Bionic chip appears to be much more powerful than Google's first in-house chip on the machine learning front as well.
The MOST powerful smartphone chip 3.0! A16 vs 8+ Gen1 vs Tensor 2 vs Exynos vs Dimensity vs Kirin!
As first highlighted by Max Weinbach on Twitter and confirmed by Notebookcheck, Apple's A15 Bionic performs better on GeekBench ML over TensorFlow Lite CPU, GPU and ML tests. While the Tensor scored 313, 1359 and 1722 in CPU, GPU and NNAPI, the A15 Bionic scored 945, 2061 and 2212 in the CPU, GPU and Core ML categories. It's a clear win for the A15 Bionic across the board, and Google Tensor isn't even close. Source: Max Weinbach at X (formerly Twitter)
While the score might be better on the A15 Bionic, that doesn't mean Tensor's ML capabilities aren't good. In fact, Google's goal with Tensor was to design a chip that works well with its own ML models. And if we follow the heterogeneous computing strategy that Google uses with Tensor, it's clear that we shouldn't completely rely on benchmarks here.
"We made sure that different subsystems inside Tensor work really well together, rather than optimizing individual elements for peak speeds. Peak CPU and GPU speeds look good in benchmarks, but they don't always reflect the real user experience," said Google Product Manager Monika Gupta at the Pixel Fall presentation keynote.
