Rumor: the Exynos 2500 might use Google's TPU for AI acceleration

2024-04-03

Samsung’s big push for AI functionality will continue next year and it will involve the Exynos 2500, according to X poster Connor/콘어/コナー. Just like Samsung uses an AMD-designed GPU in its flagship chips, it will also use a Google-designed TPU.

The TPU or Tensor Processing Unit has been a part of Pixel phones for several years and is integrated into the Tensor chipsets that are co-developed with Samsung. It is one of several types of AI accelerators and is in heavy use by Google products – you can even rent bigger versions of the TPU from Google Cloud if you are working with AI.

This will give phones using the Exynos 2500 – Galaxy S25 series for starters – access to the Google machine learning APIs that can run models on the device. The current iteration of Galaxy AI is mostly cloud-powered, since even though Samsung beefed up the AI hardware on the Exynos 2400, it’s still not enough.

Based on Connor’s post, the new flagship chipset will also include two Neural Processing Units – G-NPU and S-NPU. The “G” stands for “general” and is a more flexible version of a fixed-function NPU. These have been a part of Exynos since the 2200.

The most recent report claims that the Galaxy S25 series will use both Exynos 2500 and Snapdragon 8 Gen 4 chips, not just Exynos as a previous rumor claimed. Qualcomm has been pushing AI acceleration too – the 8 Gen 3 can run models with up to 10 billion parameters – but its AI Engine is different from the TPU. Maybe the differences will be smoothed over by Android’s machine learning API’s?

Leaked info suggests that the Exynos 2500 will be fabbed on Samsung’s second generation 3nm node (3nm GAP/SF3). It will feature a Cortex-X5 at 3.2GHz, three A730 cores at 2.3-2.5GHz, two more A730 at a lower clock speed and four A520 cores. It will use an Xclipse 950 GPU, but no details are available on it.

For comparison, the Tensor G4 (which will likely be the source of the TPU design) will be fabbed on Samsung’s 4LPP+ node (after talks with TSMC fell through). That means the Exynos 2500 should be faster and more efficient, even in AI tasks (which was Google’s stronghold until now).

Source | Via