Why the new GPU won’t make the Google Pixel 10 a gaming beast

We’ve barely finished processing arrivals. Pixel 9 series, but our latest leaks already predict Pixel 10 and the new generation of Google Tensor G5 CPU. While the upcoming chip’s processor performance appears to be taking a step forward, leaked specs suggest big changes in the graphics department with the introduction of Imagination Technologies’ DXT architecture, specifically the dual-core DXT-48-1536 clocked at 1.1GHz.

Imagination Technologies may not be a familiar name to you in today’s mobile chipset market. You’ll find its GPUs in unusual mid-range designs, like the 2022 MediaTek Dimensity 930, but you’ll likely remember it from earlier iPhone silicon. The Imagination PowerVR architecture was used in models prior to the A10 Fusion, before Apple licensed its intellectual property to more custom GPUs in-house. The return to flagship silicon with the Google Tensor G5 is an exciting development.

To be honest, Google’s Tensor series is underwhelming in terms of graphics, falling at least two generations behind the industry’s fastest in terms of performance. Additionally, the company has been slow to introduce new GPU designs and continues to shy away from supporting ray tracing, a niche feature we now expect from a flagship mobile GPU. This looks set to change, at least to some extent, with the arrival of the Tensor G5 and DXT-48 GPU.

I’m not going to get hung up on specific performance metrics; It’s too early for that, and the DXT architecture is an unknown quantity when it comes to mobile benchmarks and games. However, the dual-core “high configuration” DXT setup boasts 1,536 FP32 FLOPs per clock, which equates to 1.69 TFLOPS at the G5’s stated 1.1GHz clock speed. While comparing TFLOPS between GPU architectures comes with caveats, there are benchmarks online for very rough comparisons.

Qualcomm – 1.7 teraflops Snapdragon X Plus GPU scores about 3200 points in Wild Life Extreme. Somewhere in this range, the Tensor G5 will be around 20-25% faster than its predecessor, at least in this test. This will be the biggest leap in Pixel graphics performance in generations, but we expect an even bigger leap if Tensor adopts Arm’s latest version. Mali-G925 architecture at 3 nm. However, it’s slower than the 2023 Snapdragon 8 Gen 2, and therefore lags significantly behind the fastest gaming phones you can buy today and future 2025 competitors that will have that kind of power. Snapdragon 8 Elite.

Google’s internal metrics, according to Android authoritysuggest that performance may jump a little higher. The graphs aren’t very clearly labeled, but indicate gains of between 35% and 60% over the G4, depending on the benchmark. That would be more important, but even Google’s data shows it lags significantly behind the latest from Apple and Qualcomm, offering performance that’s still not as fast as 2023’s leading silicon.

The expected Tensor G5 GPU won’t see it vying for the performance crown then, but sustained performance could still be an interesting point of comparison. Fortunately, the DXT architecture has some interesting features that close the gap with the competition.

Ray tracing remains optional for DXT, as was the case with Arm’s Mali/Immortalis split. Google chooses the smallest radial acceleration cluster (RAC) block configuration it can (DXT-48-1536-0.5RT2), with half the RAC in each core. Again, the G5 isn’t aiming for monstrous performance.

However, Imagination offers the industry’s only Level 4 ray tracing implementation that may punch above its weight. Imagination enables full ALU offloading (freeing up GPU resources), BVH processing (much faster intersection calculations), and ray coherence sorting (batch processing of adjacent rays) in hardware, thereby accelerating ray tracing performance. Neither Arm’s Immortalis nor Qualcomm’s Adreno support BVH or Ray Coherency in hardware. However, we have yet to test Imagination’s long-touted ray tracing claims, so I won’t raise my expectations too high.

There are a few more interesting tidbits in the Imagination DXT whitepaper. The architecture supports up to 2×4 and 4×4 fragment shading speeds (also known as variable-rate shading), which you’ll already find in the current Tensor’s Arm Mali-G715 and other high-end platforms. There is also standard ASTC texture compression, but with HDR support. The key takeaway is that this is a GPU architecture that is very competitive in terms of features.

We also know that the new GPU supports virtualization, which the current Tensor chips don’t have. This allows accelerated graphics to be used in a virtual machine, potentially allowing Google to implement one of its many virtualization based features on Pixel 10. Perhaps the new features are one of the reasons for changing the GPU supplier?

One of the most interesting aspects of the Imagination GPU architecture is the 128-bit superscalar ALUs combined with a decentralized multi-core approach to the GPU cores. The former means that arithmetic logic units process multiple pieces of 32- or 16-bit data simultaneously, with the added benefit that wide registers are easily adaptable to a range of high- and low-precision data types.

This is a different approach than other mobile GPU architectures, where you typically find dedicated 32- and 16-bit ALUs running concurrently with the most common graphics data sizes, with smaller data sizes optionally supported in these machine learning ALUs. The traditional setting is good for graphics and quite good for lower-bit machine learning workloads. However, it cannot use SIMD for larger data types, which can benefit memory bandwidth and cache resources, which are always in short supply on mobile GPUs.

Combined with two GPU cores that operate independently, this also means potentially higher performance and/or lower power consumption for graphics processing and compute workloads due to the efficiency of parallel processing. In other words, you can force cores to run one or more workloads as quickly as possible, or disable a core to save power.

Additional efficiency savings when working with graphics and/or machine learning may have caught Google’s attention. However, these cores cannot share internal resources, which can lead to bottlenecks or underutilization compared to a unified shader architecture (like Mali), so there are risks involved. We’ll just have to wait and see how he behaves.

Speaking of AI, I crunched some numbers given in Google’s internal docs and calculated that the DXT-45 is about 5% faster at performing FMA operations than the G-715, which isn’t that much. However, it’s possible that the larger 128-bit register means the DXT will still be able to do more operations via SIMD and/or better use of bandwidth. It will be interesting to see if Google uses the GPU for AI workloads, especially since its TPU is only expected to see a 14% boost in the next generation.

However, I’m not sure if compute or gaming performance is the reason for the switch – DXT doesn’t look like it’ll be able to beat the competition here. The real reason for the replacement probably lies somewhere in the balance between IP cost, power efficiency and feature set offered. Either way, Google seems to have decided that Imagination Technologies is the best option going forward.

Unfortunately, those who were hoping that the move to a new GPU would propel the Tensor G5 and Pixel 10 up the graphics leaderboards will be disappointed. While a modest gain of 25% to a potentially much larger 60% is very welcome, the chip will still be two years behind the leaders. Worse, its GPU appears to have stalled again with the Tensor G6, leaving the Pixel 11 even further behind the pace.

However, as part of the transition, Tensor G5 will receive several tools. Ray tracing support, a different architecture for GPU-intensive tasks, and GPU virtualization mean that the Pixel 11 definitely won’t be short on features and will be an upgrade for gamers. Perhaps this will help Google suggest some more interesting Exclusive Pixel Features who keep the series in the spotlight.

But this is getting ahead of ourselves; The Pixel 10 is almost a year away from launch, and the competition is already moving forward. While Google is preparing some CPU and GPU changes for the Tensor G5, the chip unfortunately still lags behind the fastest processors in the business. Will Google’s leadership in artificial intelligence be enough to keep competitors at bay? I’m increasingly worried that this won’t happen.