Google Tensor G4: 45 Tokens/s for 3.5B Parameters

On August 13, Google officially introduced its latest Pixel 9 series, including the Pixel 9 Pro Fold, Pixel 9, Pixel 9 Pro, and Pixel 9 Pro XL models. These new smartphones are equipped with Google's latest Tensor G4 processor and advanced AI capabilities powered by Gemini AI, with starting prices set at $799.

Although detailed specifications of the Tensor G4 processor were not fully disclosed, reports indicate it is built using Samsung's 4nm process technology (4LPP+). The processor features a 3.1GHz Arm Cortex-X4 core, three 2.6GHz Arm Cortex-A720 cores, and four 1.92GHz Arm Cortex-A520 cores. The GPU remains the Arm Mali-G715 MC7 from the previous generation, with a frequency boost to 940Hz but no ray-tracing support. It also includes a third-generation Tensor Processing Unit (TPU) and has been upgraded with the Samsung Exynos Modem 5400, which supports 4G/5G, Wi-Fi 7, Bluetooth 5.x, and introduces satellite connectivity. The Tensor G4 supports LPDDR5X memory and can decode various video formats including H.264, H.265, VP9, and AV1.

The Tensor G4 processor, based on current information, may not directly compete with high-end processors such as the Qualcomm Snapdragon 8 Gen 3 or MediaTek Dimensity 9300. Its performance may fall short compared to last-generation flagship processors. Even when compared to its predecessor, the Tensor G3, the Tensor G4 shows only modest improvements, lacking the significant performance gains seen in previous generations.

Despite these limitations, the Pixel 9 series benefits from the new third-generation Tensor Processing Unit and the Gemini Nano model, which enhances computational photography and supports generative AI functions at 45 tokens per second. However, compared to other flagship smartphones, the generative AI performance of the Pixel 9 series is relatively modest.

For instance, the Qualcomm Snapdragon 8 Gen 3 can handle models with up to 10 billion parameters, achieving 15 tokens per second. The MediaTek Dimensity 9300 supports models with up to 13 billion parameters, with a generation rate of 20 tokens per second for models with 7 billion parameters. Although these processors have slower generative AI speeds, they support models with parameters more than double those of the Gemini Nano, offering superior generative AI capabilities.

The Pixel 9 series also features significant memory upgrades, with the Pixel 9 offering 12GB of RAM and the Pixel 9 Pro equipped with 16GB. This increase in memory supports the operation of large models.

Google has allocated a portion of the RAM specifically for running the Gemini model, ensuring a smoother AI experience by preventing memory usage by other applications. However, the exact amount of RAM reserved for this purpose has not been disclosed.

Additionally, the Tensor G4 retains many of the internal components from the Tensor G3, including the AI-focused TPU, GXP digital signal processor for camera and audio tasks, BigWave media codec block, and Titan M2 security chip. These components remain unchanged, indicating similar functionality to previous models.

A notable improvement is the new Samsung Exynos Modem 5400, which addresses previous connectivity and battery efficiency issues. Reportedly, this modem is 50% more efficient than the Exynos Modem 5300 found in the Pixel 8 series. It also supports 3GPP Rel. 17, including 5G Non-Terrestrial Networks (NTN), enabling the Pixel 9 series to utilize new satellite communication features for sending SOS messages.