Google’s Tensor DLA Lags In Area, Speed

Among flagship-smartphone processors, Google’s first custom design has the largest AI engine but the worst AI performance. In contrast, the Dimensity 9000’s AI engine is the most efficient.

Linley Gwennap

Known for its AI prowess, Google made the deep-learning accelerator (DLA) a showpiece in its custom Tensor processor for the Pixel 6 smartphone. At its release, the Tensor chip led in some AI benchmarks, but next-generation flagship processors from Qualcomm and MediaTek quickly surpassed it. Our analysis of AI benchmarks reveals the Apple and Samsung processors also outperform Google’s.

We additionally analyzed TechInsights teardown and floorplan reports on these chips, including Tensor, the Apple A15, the Dimensity 9000, the Snapdragon 8 Gen 1, and the Exynos 2200 as well as predecessor parts. Of these, Tensor’s 16-core design required more die area than any competing DLA. Thus, Google achieved the dubious distinction of getting the least from the most.

One challenge for all vendors is rapidly rising performance: MediaTek, Qualcomm, and Samsung all doubled throughput over their previous generation. These gains account for both hardware and software upgrades, unlike TOPS (trillions of operations per second), which is a hardware-only metric. Perhaps for this reason, most vendors have stopped disclosing TOPS ratings in the most recent processor generation.

AI benchmarking remains fragmented. ETH Zurich’s generically named AI-Benchmark offers extensive information on a variety of processors, but it supports only Android’s NNAPI software. Geekbench recently delivered its first AI test, Geekbench ML, which supports both NNAPI and Apple’s CoreML interface, providing a multiplatform measurement. Its database of results, however, is smaller than AI-Benchmark’s. MLPerf has a Mobile test suite, but few vendors have followed its complex submission process. We also looked at ETH’s Burnout test, mainly to assess the power consumed by Android phones.