Google Pixel 8's Tensor G3 GPU tests show weak performance but decent efficiency

Peter, 12 October 2023

Google teamed up with Samsung once again to build the Tensor G3, the chipset that powers the new Pixel 8 and Pixel 8 Pro. It comes out of Samsung’s 4nm foundries and features a 9-core CPU (1x Cortex-X3 + 4x A715 + 4x A510) and an ARM Mali-G715 GPU.

Google hasn’t revealed specifics about this GPU, but it seems to have 7 compute units, the same as last year’s Tensor G2 (that one was a Mali-G710). How efficient is this GPU? Golden Reviewer has run some tests to find out.

Starting with 3DMark Wildlife Extreme, the chipsets are rated based on score divided by average power (so higher is better). The G3 doesn’t do so well – it’s about as fast as an Apple A14 and is slower than the Snapdragon 8 Gen 1.

Google Pixel 8's Tensor G3 GPU tests show weak performance but decent efficiency

The chipset to keep an eye on in these tests is actually MediaTek’s Dimensity 9200. It also uses the Mali-G715 (with 11 compute units) but has been fabbed on a 4nm node from TSMC. The Dimensity outperforms the Tensor by a wide margin, however, it is nearly equal in terms of efficiency.

Moving over to GFXBench Aztec (Vulkan/Metal back end, 1440p offscreen), the Dimensity once again schools the Tensor and offers nearly double the frame rate, but draws nearly double the power too, so it’s basically the same efficiency-wise. Somewhat disappointingly the Tensor G3 is only a small upgrade over the G2 in this test (the gap in Wildlife was much more significant).

Google Pixel 8's Tensor G3 GPU tests show weak performance but decent efficiency

Finally, another GFXBench test (v3.1) shows the gap between the Google chip and the MediaTek one is smaller at 1080p. Still, accounting for power draw the two chipsets are about the same.

Google Pixel 8's Tensor G3 GPU tests show weak performance but decent efficiency

According to Golden Reviewer, Samsung’s 4nm LPP node is 1.5-2 generations behind the 4nm TSMC node. In the more demanding tests the Tensor G3 is just not competitive with current flagship chips, never mind the ones that are just behind the corner – the Snapdragon 8 Gen 3 and Dimensity 9300 should be unveiled before the end of this year.

The node disadvantage doesn’t bode well for Samsung’s own Exynos 2400, though that one has a completely different GPU – one based on RDNA 3 tech from AMD, so an architectural advantage could keep it in the game.

Back to Google, the Tensor G4 will likely be built at Samsung's foundry once again as Google still needs the expertise of the Exynos team. However, rumor has it that a Google is working on an in-house design that won't depend on Samsung. That is expected in 2025.

Source


Related

Reader comments

The 4LPP node (or worse, the older 4LPE node) may have been mistaken here for the newer 4LPP+ process, which the Tensor G3 uses if I remember correctly. Edit: no, I was wrong. Tensor G3 is built on the standard 4LPP node which improves on power ...

  • Yes Ted
  • 17 Oct 2023
  • K0E

Top GPU only matters to an extremely small % of users and is still good enough to run every game out there without issue (except high demanding emulation). All that matters to the vast majority is UX which CPUs have well outpaced in terms of power re...

Only correct thing is the last sentence. And not in the way you meant it.

Popular articles

More

Popular devices

Electric Vehicles

More