Submit a review

Tensor G3 in the Google Pixel 8: Leak reveals various details

Thanks to a source within Google itself, there are new details on the Google Pixel 8 series and the Tensor G3 chip that will power them.
Google Pixel 8

Two years ago, Google introduced the Tensor processor, its first custom chipset for smartphones. Thanks to a long-term partnership with Samsung’s semiconductor department and own development experience, the second-generation Tensor chip is now powering the Pixel 7 series. Even if the project is sometimes criticized for putting absolute top performance in the background in favor of AI performance, the success of the latest Pixel models cannot be denied.

Tensor has also given Google the ability to leverage its AI expertise and create brand new experiences that would otherwise be impossible, which have become core to the Pixel series’ identity. Thanks to a source within Google itself, we’re now getting fresh insights into the upcoming Google Pixel 8 series, as well as the chipset that will power them – the Tensor G3, codenamed zuma.

Tensor G3 has more modern CPUs

The Tensor G2 was a rather uninspiring chipset in terms of CPU performance. At release, all cores were already two generations behind the competition. Additionally, the only real change from the first-gen chip was a mid-cluster upgrade from the old Cortex-A76 cores to more matching Cortex-A78 cores. The chip retained the unusual 4+2+2 core layout where most other chip makers used a 4+3+1 layout with a single large prime core.

With the Tensor G3, Google is finally packing more up-to-date cores into the chip. The entire CPU block has been restructured to use 2022 ARMv9 cores. The core layout has also been changed. The unusual 4+2+2 setup is gone and instead the Tensor G3 will have nine CPU cores. There are four small Cortex-A510s, four Cortex-A715s and a single Cortex-X3. At the same time, frequencies will increase compared to previous generations. Thus, the performance should increase significantly and the Tensor G3 should be on par with other flagship processors of 2022. Chipsets with the newly announced ARMv9.2 cores, on the other hand, are not within reach.

Tensor Clock Rates

CPUTensor G3 (zuma)Tensor G2 (gs201)Tensor (gs101)
Prime Cores1x Cortex-X3 @ 3.0GHz2x Cortex-X1 @ 2.85GHz2x Cortex-X1 @ 2.8 GHz
Mid Cores4x Cortex-A715 @ 2.45GHz2x Cortex-A78 @ 2.3GHz2x Cortex-A76 @ 2.25GHz
Little Cores4x Cortex-A510 @ 2.15GHz4x Cortex-A55 @ 1.8GHz4x Cortex-A55 @ 1.8GHz

The move to ARMv9 also allows Google to implement new security technologies – the Pixel 8 will feature ARM’s Memory Tagging Extensions (MTE), which can prevent some memory-based attacks. Other smartphones already support MTE on the hardware side, but have not yet activated the function in Android. The Pixel 8’s bootloader now appears to be the first to implement this interface.

The most important change in ARMv9 is of course the switch to pure 64-bit code. Tensor G2 devices like the Google Pixel 7 and Pixel 7 Pro had already dropped support for older 32-bit apps, but still keep the 32-bit libraries on board. Quite apart from the 32-bit capable cores, which changes with the Google Pixel 8. The phone ships with 64-bit software only, but it’s unclear if the Cortex-A510 cores are configured with AArch32 support.

Integrated ray tracing graphics

Graphics have always been at the heart of Google’s Tensor processors, even if the latest Tensor G2 didn’t set any benchmark records. The original Tensor’s original 20-core Mali-G78 configuration (out of a maximum of 24 cores) outperformed Qualcomm’s Snapdragon 888 and Samsung’s Exynos 2100, but was quickly overtaken by newer models. Still, powerful graphics chips are useful for neural network applications that run more efficiently on a GPU than on a CPU.

Although Google has upgraded to a newer Mali-G710, Tensor G2 benchmarks showed that the seven-core setup only delivered better sustained load and no noticeable increase in peak graphics performance. The Tensor G3 in the Pixel 8 will fix this with the upgrade to the ARM Mali-G715 GPU.

GPUTensor G3 (zuma)Tensor G2 (gs201)Tensor (gs101)
GPU Core ModelMali-G715 (Immortalis)Mali-G710Mali-G78
Core Count10720
Frequency (shaders)890MHz848MHz848MHz

Although the source could not specify the exact number of cores, various hardware configuration details point to an MP10 setup (ten cores). That would make the GPU the Immortalis variant of the G715, complete with ray tracing capabilities.

First smartphone chip with AV1 coding

The first generation Google Tensor used a hybrid architecture for video acceleration. A generic Samsung Multi-Function Codec (MFC) IP block was used, the same as the Exynos chips, but explicitly left out the AV1 support. This is where Google’s custom “BigOcean” hardware video decoder block came into play. “BigOcean” supports up to 4K60fps AV1 video decoding. The Tensor G2 left the hardware block mostly unchanged and retained the same decoding capabilities.

The Tensor G3 finally updates the video block. First, the MFC block now supports 8K30 video decoding/encoding in H.264 and HEVC. It’s important to note, however, that a special internal version of the Google Camera used to test the Pixel 8 series doesn’t currently support 8K video recording. Pixel phones already struggle with thermals when shooting 4K. Not to mention how quickly it fills up memory.

CPUTensor G3 (zuma)Tensor (gs101) | Tensor G2 (gs201)
H.264 decode8K30 / 4K120 / 720p240 (MFC)4K120 / 720p240 (MFC)
H.264 encode8K30 / 4K120 / 720p240 (MFC)4K120 / 720p240 (MFC)
HEVC decode8K30 / 4K120 / 720p240 (MFC)4K120 / 720p240 (MFC)
HEVC encode8K30 / 4K120 / 720p240 (MFC)4K120 / 720p240 (MFC)
AV1 decode4K60 / 1080p120 (BigWave)4K60 / 1080p120 (BigOcean)
AV1 encode4K30 / 720p240 (BigWave)

More importantly, Google has now evolved the “BigOcean” block into “BigWave”. While the video decoding capabilities remain the same (up to 4K60fps AV1 video), the block now also supports AV1 encoding at up to 4K30fps. This makes Google the first smartphone brand to ship an AV1 encoder in a mobile device. It will be interesting to see how it is used as the 30fps limit is not ideal for video recording.

AI

Traditionally, Tensor processors have undoubtedly focused on AI performance. After distilling its EdgeTPU Server ML accelerators onto the Pixel 4’s Pixel Neural Core, Google’s first-gen Tensor already shipped with an integrated TPU codenamed “Abrolhos.” With a clock speed of 1.0 GHz, it delivered excellent performance especially in NLP (Natural Language Processing) tasks.

The Tensor G2 came with a TPU called “Janeiro” that also runs at 1.0 GHz. According to Google, Janeiro is up to 60% faster than the original chip for camera and voice tasks. The Tensor G3 also brings a new version of the TPU codenamed “Rio” and clocked at 1.1 GHz. There is no concrete data on Rio’s performance yet, but it should be a significant upgrade.

More Tensor G3 improvements for the Google Pixel 8 series

The Tensor G2 also introduced a new element that wasn’t much discussed – Google’s custom digital signal processor (DSP) “Aurora”, also called GXP. DSPs are specialized processors for tasks like image processing, and that’s exactly how Google uses them. GXP replaces the GPU in many common image processing steps, such as B. Blur and local tone mapping. This makes these general operations faster and more efficient.

However, the Tensor G2 still has a first-generation GXP (codename “Amalthea”) in a 4-core configuration with 512 KB of memory per core and 975 MHz clock speed. The Tensor G3 features a brand new second generation GXP (codename “callisto”) in a similar 4-core configuration with 512 KB per core and a modest frequency increase to 1,065 MHz.

Faster UFS storage

The Tensor G3 also brings a new version of Samsung’s UFS controller, which now supports UFS 4.0 storage. UFS 4.0 is a major upgrade over UFS 3.1, doubling theoretical speeds while delivering 50% better efficiency. Other flagship smartphones like the Samsung Galaxy S23 Ultra already have UFS 4.0 storage. With the improved controller, the Google Pixel 8 catches up in this respect and closes the gap to the competition.

No major modem upgrades

One of the original Tensor’s biggest flaws was its underpowered Samsung Exynos 5123 modem. It lagged behind other vendors in terms of performance and supported standards, and it also had significant power consumption and temperature issues. Not to mention the initial stability issues, although these have been greatly reduced with software updates.

The Tensor G2 switched to the Exynos 5300 modem. It brought performance and efficiency improvements, but mostly didn’t solve the heat and power consumption issues. The Tensor G3 is rumored to continue using the same modem, although it’s said to be a slightly different variant.

Tensor G3 will power the Google Pixel 8

That’s all that’s known so far about Google’s upcoming chip. Tensor has given Google more control over the direction of its smartphone brand. At the same time, the Google Pixel smartphones offer some unique features that will also be crucial to the Pixel 8 series performance.

Unlike the Tensor G2, which was a minor update, the Tensor G3 looks set to be a major upgrade. Google wants to become more competitive in processing all apps, and not just AI tasks. With the upcoming CPU and GPU upgrades, this plan could work out.

Source

Achim Maier
We will be happy to hear your thoughts

Leave a Reply

S24
Register New Account
Compare items
  • Total (0)
Compare
0