We have barely finished processing the arrival of the Pixel 9 series, but our latest leaks already anticipate the Pixel 10 and the new generation of Google Tensor G5 processor. While the upcoming chip’s CPU performance appears to take a step forward, the leaked specs suggest a bigger change in the graphics department with the adoption of Imagination Technologies’ DXT architecture, specifically a DXT-48- 1536 dual-core clocked at 1.1 GHz.

Imagination Technologies may not be a name you’re very familiar with in today’s mobile chipset market. You’ll find its GPUs in a weird midrange design, like MediaTek’s Dimensity 930 from 2022, but you’re more likely to remember it from the previous iPhone’s silicon. Imagination’s PowerVR architecture powered models up to the A10 Fusion before Apple licensed its IP for more bespoke internal GPUs. The return to flagship silicon with Google’s Tensor G5 is an exciting development.

To be frank, Google’s Tensor series is disappointing in the graphics department, languishing at least two generations behind the industry’s fastest in terms of performance. It has also been slow to adopt new GPU designs and continues to avoid ray tracing support, a niche feature but one we now expect in a flagship mobile GPU. This should change, at least somewhat, with the Tensor G5 and the DXT-48 GPU.

I’m not going to focus on specific performance numbers; it’s far too early for that, and the DXT architecture is an unknown quantity when it comes to benchmarks and mobile titles. Still, a “high configuration” DXT configuration with two SPU cores offers 1,536 FP32 FLOPs per clock, putting it at 1.69 TFLOPs at the G5’s reported clock speed of 1.1 GHz. Although comparing TFLOPS on GPU architectures is fraught with pitfalls, there are benchmark figures circulating online for a very rough comparison.

Qualcomm’s TFLOP 1.7 Snapdragon XPlus GPU scores around 3200 in Wild Life Extreme. Somewhere in that range, the Tensor G5 would be around 20-25% faster than its predecessor, at least in this test. This would be the most significant step forward in Pixel graphics performance in generations, but we’d expect an even bigger jump if Tensor adopted Arm’s latest version. Mali-G925 architecture on 3nm. Regardless, it runs slower than 2023’s Snapdragon 8 Gen 2 and, therefore, well below the pace of the fastest gaming phones you can buy today and upcoming 2025 rivals which pack the power station. Snapdragon 8 Elite.

Google’s internal figures, seen by Android Authoritysuggest that performance could increase a bit more. The graphs are not well labeled but indicate a 35-60% gain over the G4, depending on the benchmark. That would be bigger, but even Google’s data shows it falls well short of the latest from Apple and Qualcomm, delivering performance that’s still not as fast as 2023’s leading silicon.

The Tensor G5’s expected GPU won’t then see it challenge for the performance crown, but sustained performance could still be an interesting point of comparison. Fortunately, the DXT architecture has some exciting features that will help narrow the gap with its competitors.

Ray tracing remains optional with DXT, as it is with Arm’s Mali/Immortalis division. Google opts for the smallest Ray Acceleration Cluster (RAC) unit configuration possible (a DXT-48-1536-0.5RT2), with half a RAC in each core. Again, the G5 isn’t aiming for beastly performance.

Still, Imagination offers what it calls the industry’s only Level 4 ray tracing implementation, which could see it punch above its weight. Imagination features full ALU offloading (freeing up GPU rendering resources), BVH processing (much faster intersection calculations), and ray coherence sorting (group processing of nearby rays) in hardware, thereby accelerating performance ray tracing. Neither Arm’s Immortalis nor Qualcomm’s Adreno support BVH or Ray Coherency at the hardware level. That said, we haven’t yet tested Imagination’s Ray Tracing claims, so I won’t set my expectations too high.

Imagination’s DXT white paper contains other interesting information. The architecture supports up to 2×4 and 4×4 fragment shading rate (aka Variable Rate Shading), which you’ll already find in the current Tensor’s Arm Mali-G715 and other platforms. high-end shapes. There is also standard ASTC texture compression, but with HDR support. The key takeaway is that this is a very competitive GPU architecture from a feature perspective.

We also know that the new GPU supports virtualization, something not found in current Tensor chips. This allows the use of accelerated graphics in a virtual machine, potentially allowing Google to bring one of its many virtualization-based features to the Pixel 10. Maybe the new features are one of the reasons why you’re switching GPU vendors?

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

This is a different approach to other mobile GPU architectures, where you’ll typically find dedicated 32- and 16-bit ALUs running simultaneously on the most common graphics data sizes, with smaller data sizes optionally supported in these ALUs for machine learning. The traditional setup is good for graphics and not bad for low-bit depth machine learning workloads either. However, it cannot leverage Single Instruction Multiple Data (SIMD) on larger data types, which can be beneficial for memory bandwidth and cache resources, which are always valuable in mobile GPUs.

Paired with two GPU cores that operate independently, this also means potentially higher performance and/or lower power consumption when running graphics and compute workloads, thanks to the efficiency of parallel processing. In other words, you can have cores contribute to a single or different workload as quickly as possible or turn off a core to save power.

Additional efficiency savings for graphics and/or machine learning workloads may have caught Google’s attention. That said, these cores can’t share internal resources, which can lead to bottlenecks or underutilization compared to a unified shader architecture (like Mali’s), so it’s not without risks . We’ll just have to wait and see how it performs.

Speaking of AI, I ran some numbers in Google’s internal documents and estimate that the DXT-45 is about 5% faster in FMA operations than the G-715, which is not a lot. However, it is possible that a larger 128-bit register means that DXT can still do more with each operation via SIMD and/or better bandwidth utilization. It will be interesting to see if Google leverages the GPU for AI workloads, especially since its TPU is only looking at a 14% gain for the next generation.

Still, I’m not convinced that compute workloads or gaming performance are the reason for the change – DXT doesn’t seem to be able to beat the competition here. The real reason for this trade-off probably lies somewhere in the balance between IP costs, 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 boost the Tensor G5 and Pixel 10 in the graphics rankings will be disappointed. Even if a modest gain of 25%, or even 60%, potentially much larger, is welcome, the chip will still lag behind the leaders by two years. Worse yet, its GPU looks set to stagnate again with the Tensor G6, leaving the Pixel 11 even further behind.

However, Tensor G5 gains some tools in the transition. Ray tracing support, a different architecture for GPU-bound tasks, and GPU virtualization mean the Pixel 11 certainly won’t be lacking in features and will be an upgrade for gamers. This might just help Google come up with some of these cool things Pixel-exclusive capabilities which keep the series in the spotlight.

But it’s getting ahead of ourselves; the Pixel 10 is almost a year away and the competition is already moving forward. While Google is preparing some CPU and GPU changes with the Tensor G5, the chip unfortunately still looks far behind the fastest processors in the industry. Will Google’s lead in AI be enough to keep its competitors at bay? I increasingly fear that this is not the case.