On paper, the Tensor G5 looks good. It features an eight-core CPU made up of one Cortex-X4 core clocked at 3.78 GHz, five Cortex-A725 cores at 3.05 GHz and two Cortex-A520 efficiency cores at 2.25 GHz. There is a fifth-generation TPU for AI tasks, an Imagination IMG DXT-48-1536 GPU at 1.10 GHz and a Samsung Exynos 5G modem. Built on TSMC’s 3nm node, it should deliver efficiency and performance in equal measure.

In the real world, it runs hot enough to melt the back of a Pixel phone so the chip begins throttling even during modest gaming or PlayStation 2 emulation, a workload that mostly hammers the CPU rather than the GPU. That points to a much deeper design flaw.

Qualcomm’s Snapdragon 8 Elite Gen 5 cleans the Tensor G5's clock in Geekbench 6 and 3DMark tests. Qualcomm’s Oryon CPU cores run at higher clock speeds, with a prime core at 4.6 GHz and performance cores at 3.62 GHz, all supported by 12 MB of L2 cache per core type. The difference is that Qualcomm designs and tunes its silicon. Google, meanwhile, is still using off-the-shelf ARM Cortex cores with little fine-tuning and poor thermal optimisation.

The GPU situation is just as awkward. Google worked with Imagination Technologies to build the DXT-48-1536, but Imagination holds the keys to the driver updates and hardware code. That leaves Google unable to push out meaningful performance optimisations or bug fixes without waiting in line.

The result is a chip that feels like an off-the-rack suit with a few custom hems. It works well enough, but it lacks the sharp fit and finish of something made to measure.

Until Google commits to a proper in-house silicon design, the Tensor range will remain a hot, throttling reminder that cost-cutting and true performance rarely share the same bed.