Tensor G5 Specs Leaked: Google May Finally Redeem Itself Next Year

Google’s custom Tensor chipsets have not been received well over the years, delivering poor efficiency and decent performance compared to the competition. Google primarily designs the Tensor in-house with some Samsung IP. Notably, Tensor is manufactured by Samsung foundry which results in lower yield and poor performance/efficiency. In our Tensor G4 benchmark testing, the chipset didn’t perform well.

Finally, that is going to change with the Tensor G5 (codename “laguna”) next year. Android Authority has revealed the complete specs of the Tensor G5 that will power the upcoming Pixel 10 series. Tensor G5 will reportedly be manufactured on TSMC’s 3nm (N3E) process node, the same as Qualcomm’s Snapdragon 8 Elite and Apple’s A18 Pro processors.

Simply moving to TSMC’s foundry will likely boost the performance and efficiency of Tensor G5. Apart from that, the report says that the Tensor G5 die size is 121 mm^2, a bit smaller than the Snapdragon 8 Elite’s 124.1 mm^2. It’s not clear if Tensor G5 includes an integrated modem.

Coming to the CPU configuration, the Tensor G5 features eight cores. However, Google is still stuck with the older Cortex-X4 core instead of going with the latest Arm Cortex-X925 core that brings a big performance jump, as we have seen in our comparison between Dimensity 9400 and Snapdragon 8 Elite. There is no information on the CPU frequency.

That said, for mid-cores, Google thankfully picked Arm’s latest 5x Cortex-A725 cores. And for efficiency cores, there are 2x Cortex-A520 cores. Looking at the on-paper spec sheet, it won’t beat the flagship chipsets in the market, but it will likely bring a significant performance uplift over the previous Tensor G4.

PowerVR Makes a Comeback With Tensor G5

As for the GPU, Google has curiously picked Imagination Technologies’ 2-core DXT-48-1536 GPU which runs at 1.1GHz and brings Ray Tracing and GPU Virtualization support. In case you are not aware, Imagination Technologies was behind the PowerVR GPU that was used in iPhones from 2007 to 2017, till Apple A10 Fusion SoC. With the A11 Bionic, Apple moved to its in-house GPU.

And for the TPU aka NPU, the Tensor G5 packs a new-gen AI accelerator that offers 14% faster performance on AI workloads. The report says it can perform 18 TOPS on the INT8 data type and 9 TOPS on the FP16 data type.

Overall, the Tensor G5 looks exciting to me because it’s finally built on TSMC’s node. That said, there is no information on the 5G modem Google will be using for Tensor G5. Will Google pick Samsung again or tap Qualcomm’s shoulder to integrate a discrete 5G modem, like Apple? Well, stay tuned with us as we dig in and find more information.