Robert Triggs / Android Authority
Google’s Tensor G5, destined for next year’s Pixel 10, might be the most awaited chip in the series yet. For the first time ever, Google is designing the chip completely internally without relying on Samsung for the bulk of the work. The expectation is this will make it more competitive, as the previous and even current Tensor models lag behind other chipset manufacturers. But will it actually happen? So far we haven’t heard anything about the specs of the new chip, but that finally changes today.
Thanks to a massive leak from Google’s gChips division, Android Authority has viewed credible documents from around the end of the Tensor G5’s design that tells us everything about Google’s next-gen chip.
More weird CPU decisions
Pixel 9’s Tensor G4 featured an upgraded CPU cluster compared to its predecessor — at least on paper because, in practice, the performance is only marginally faster — just 6% quicker in our multi-core tests. This is because Google dropped a mid-cluster core, probably in an attempt to reduce power usage.
The Tensor G5 upgrades the CPU cluster again, but not in a way you’d probably expect. Google has decided to keep the same single Arm Cortex-X4 primary core, which is an interesting choice given the new Cortex-X925 promises some big improvements. It also decided to shift the core clusters around again: the mid cluster now has five Cortex-A725 cores, instead of three Cortex-A720 on Tensor G4, and the little cluster has been shrunk accordingly to only two Cortex-A520 cores. I’ve compiled the specs below:
Tensor G3 (“zuma”) | Tensor G4 (“zumapro”) | Tensor G5 (“laguna”) | |
---|---|---|---|
Big cluster |
Tensor G3 (“zuma”)
1x Arm Cortex-X3 |
Tensor G4 (“zumapro”)
1x Arm Cortex-X4 |
Tensor G5 (“laguna”)
1x Arm Cortex-X4 |
Mid cluster |
Tensor G3 (“zuma”)
4x Arm Cortex-A715 |
Tensor G4 (“zumapro”)
3x Arm Cortex-A720 |
Tensor G5 (“laguna”)
5x Arm Cortex-A725 |
Little cluster |
Tensor G3 (“zuma”)
4x Arm Cortex-A510 |
Tensor G4 (“zumapro”)
4x Arm Cortex-A520 |
Tensor G5 (“laguna”)
2x Arm Cortex-A520 |
Obviously, it’s hard to predict performance based on specs alone, but a change like this should give the Tensor G5 at least a decent multi-core performance improvement. It is disappointing, however, to see the same Cortex-X4 used again.
A brand-new GPU
One area that particularly surprised me about the Tensor G5 while reviewing the documents was the GPU. There aren’t many vendors providing GPU IP, and Google has been using Arm Mali for all its past Tensor chips, so why would that change now? Well, I don’t know, but it did. The Tensor G5 comes with a GPU from Imagination Technologies (or IMG) — the DXT-48-1536 running at 1.1 GHz.
Unfortunately, we don’t know much about the new GPU except for two interesting details: first, it will come with ray tracing support, which is new to Google chips, which typically skip such “gaming” features. The second and perhaps just as important detail is the support for GPU Virtualization, allowing the use of accelerated graphics in a virtual machine. Google has been working on various virtualization-based features for a while, so it makes sense this was included.
I’ve compiled the new GPU specs in the table below:
Tensor G3 (“zuma”) | Tensor G4 (“zumapro”) | Tensor G5 (“laguna”) | |
---|---|---|---|
GPU |
Tensor G3 (“zuma”)
Arm Mali-G715 (7 cores) |
Tensor G4 (“zumapro”)
Arm Mali-G715 (7 cores) |
Tensor G5 (“laguna”)
IMG DXT (2 cores) |
Frequency |
Tensor G3 (“zuma”)
890 MHz |
Tensor G4 (“zumapro”)
940 MHz |
Tensor G5 (“laguna”)
1100 MHz |
Ray tracing |
Tensor G3 (“zuma”)
Not supported |
Tensor G4 (“zumapro”)
Not supported |
Tensor G5 (“laguna”)
Supported |
GPU Virtualization |
Tensor G3 (“zuma”)
Not supported |
Tensor G4 (“zumapro”)
Not supported |
Tensor G5 (“laguna”)
Supported |
Modest improvements for AI
C. Scott Brown / Android Authority
Google’s phones have always had great AI features, and AI is obviously one of the main reasons why Tensor chips even exist. By including custom-designed TPUs, Google can achieve much more impressive experiences than would otherwise be possible with an off-the-shelf chip.
The Tensor G5 comes with a modestly faster TPU compared to the Tensor G4 (which itself was exactly the same as the Tensor G3). The TOPS (Trillions of Operations Per Second) value is almost 40% larger, but that doesn’t translate well to real-world performance. Google’s internal benchmarks suggest the new TPU is only 14% faster. The improved TPU also comes with a few new features for Google’s developers, such as small embedded RISC-V cores to allow running operations that are not implemented in hardware, as well as support for on-device training. I’ve compiled the new TPU specs below:
Tensor G3 (“zuma”) | Tensor G4 (“zumapro”) | Tensor G5 (“laguna”) | |
---|---|---|---|
TOPS |
Tensor G3 (“zuma”)
13 / 6.5 TOPS (INT8/FP16) |
Tensor G4 (“zumapro”)
13 / 6.5 TOPS (INT8/FP16) |
Tensor G5 (“laguna”)
18 / 9 TOPS (INT8/FP16) |
Performance improvement over previous generation |
Tensor G3 (“zuma”)
+65% |
Tensor G4 (“zumapro”)
0 |
Tensor G5 (“laguna”)
+14% |
As we’ve already revealed, the Tensor G5 is built on TSMC’s 3 nm-class N3E process node, similar to Apple’s A18 Pro, for example. Interestingly the die size is 121 mm^2, whereas the Apple A18 Pro’s is only 105 mm^2, making the G5 a noticeable larger chip.
While the specs of Google’s upcoming Tensor G5 don’t look super impressive on paper, it’s worth remembering that it’s still a chip that will be in a Pixel phone, where software is arguably more important than the hardware. It will be interesting to see how Google utilizes the new hardware to bring new features to the Pixel 10 series.
What do you think about the leaked Tensor G5 specs?
658 votes