Why a brand new GPU will not make the Google Pixel 10 a gaming beast

We’re barely by way of processing the arrival of the Pixel 9 sequence, but our newest leaks are already waiting for the Pixel 10 and Google’s next-gen Tensor G5 processor. Whereas the longer term chip’s CPU efficiency seems to be taking a sideward step, leaked specs recommend an even bigger change within the graphics division with the adoption of Creativeness Applied sciences’ DXT structure — particularly a two-core DXT-48-1536 clocked at 1.1GHz.

Creativeness Applied sciences won’t be a reputation you’re tremendous acquainted with in at this time’s cellular chipset market. You’ll discover its GPUs within the odd mid-range design like 2022’s MediaTek’s Dimensity 930, however you’re extra prone to bear in mind it from earlier iPhone silicon. Creativeness’s PowerVR structure powered fashions as much as the A10 Fusion earlier than Apple licensed its IP for extra bespoke in-house GPUs. A return to flagship silicon with Google’s Tensor G5 is an thrilling growth.

To be blunt, Google’s Tensor sequence underwhelms within the graphics division, languishing not less than two generations behind the quickest within the enterprise by way of efficiency. It’s additionally been sluggish to undertake new GPU designs and continues to dodge ray-tracing help, a distinct segment function however one which we now count on in a flagship-tier cellular GPU. That appears set to vary, not less than considerably, with the Tensor G5 and the DXT-48 GPU.

I’m not going to fixate on particular efficiency numbers; it’s far too early for that, and the DXT structure is an unknown amount in terms of cellular benchmarks and titles. Nonetheless, a two-SPU core “Excessive Configuration” DXT setup boasts 1,536 FP32 FLOPs per clock, placing it at 1.69 TFLOPs on the G5’s reported 1.1GHz clock velocity. Whereas evaluating TFLOPS throughout GPU architectures is fraught with caveats, there are benchmark numbers floating about on-line for a really tough comparability.

Qualcomm’s 1.7 TFLOP Snapdragon X Plus GPU scores round 3,200 in Wild Life Excessive. Someplace in that ballpark would make the Tensor G5 about 20% to 25% sooner than its predecessor, not less than on this check. That might be essentially the most important leap within the Pixel’s graphic efficiency for generations, however we’d count on a fair bigger leap if Tensor adopted Arm’s newest Mali-G925 structure on 3nm. Regardless, that works out slower than 2023’s Snapdragon 8 Gen 2 and, subsequently, nicely off the tempo of the quickest gaming telephones you should purchase at this time and upcoming 2025 rivals packing the powerhouse Snapdragon 8 Elite.

Google’s inner figures, as seen by Android Authority, recommend efficiency might leap a bit larger. The graphs aren’t nicely labeled however level to a 35% to 60% acquire over the G4, relying on the benchmark. That might be extra important, however even Google’s knowledge reveals it touchdown nicely wanting Apple and Qualcomm’s newest, providing efficiency that’s nonetheless not as zippy as main 2023 silicon.

The Tensor G5’s anticipated GPU gained’t see it contesting the efficiency crown then, however sustained efficiency would possibly nonetheless make for an attention-grabbing comparability level. Fortunately, the DXT structure sports activities some attention-grabbing options that may shut the hole on its rivals.

Ray tracing stays non-obligatory with DXT, as it’s with Arm’s Mali/Immortalis cut up. Google is choosing the smallest Ray Acceleration Cluster (RAC) unit configuration it might probably (a DXT-48-1536-0.5RT2), with a half RAC in every core. Once more, the G5 shouldn’t be aiming for beastly efficiency.

Nonetheless, Creativeness sports activities what it calls the {industry}’s solely Degree 4 ray tracing implementation, which could see it punch above its weight. Creativeness sports activities full ALU offloading (releasing up GPU rendering assets), BVH processing (a lot sooner intersection calculations), and Ray Coherency Kind (group processing of close by rays) in {hardware}, thereby accelerating ray tracing efficiency. Neither Arm’s Immortalis nor Qualcomm’s Adreno helps BVH or Ray Coherency in {hardware}. That stated, we’re but to check Creativeness’s long-touted ray-tracing claims, so I gained’t set my expectations too excessive.

Creativeness’s DXT white paper accommodates another attention-grabbing tidbits. The structure helps as much as 2×4 and 4×4 Fragment Shading Fee (aka Variable Fee Shading), which you’ll already discover within the present Tensor’s Arm Mali-G715 and different high-end platforms. There’s additionally industry-standard ASTC texture compression however with HDR help. The important thing takeaway is that it is a GPU structure that’s very aggressive from a function standpoint.

We additionally know the brand new GPU helps virtualization, which isn’t present in present Tensor chips. This permits for the usage of accelerated graphics in a digital machine, doubtlessly permitting Google to carry one in all its quite a few virtualization-based options to the Pixel 10. Maybe, new options are one of many causes for switching GPU distributors?

One of many extra attention-grabbing facets of Creativeness’s GPU structure is its 128-bit Superscalar ALUs, mixed with a Decentralised Multi-Core method to GPU cores. The previous means the arithmetic logic items course of a number of items of 32- or 16-bit knowledge without delay, with the added perk that extensive registers are extremely adaptable for a spread of high- and low-precision knowledge varieties.

It is a completely different method from different cellular GPU architectures, the place you’ll usually discover devoted 32- and 16-bit ALUs working concurrently on the most typical graphics knowledge sizes, with smaller knowledge sizes optionally supported inside these ALUs for machine studying. The standard setup is sweet for graphics and never unhealthy for decrease bit-depth machine studying workloads both. Nonetheless, it might probably’t leverage single instruction a number of knowledge (SIMD) on bigger knowledge varieties, which could be helpful for reminiscence bandwidth and cache assets, that are at all times at a premium in cellular GPUs.

Paired with two GPU cores that work independently additionally means doubtlessly larger efficiency and/or decrease energy consumption when crunching by way of graphics and compute workloads, due to parallel processing efficiencies. In different phrases, you may have cores contribute to a single or completely different workload as quick as potential or energy down a core to save lots of vitality.

Further effectivity financial savings for graphics and/or machine studying workloads could have caught Google’s eye. That stated, these cores can’t share inner assets, which might result in bottlenecks or underutilization in comparison with a unified shader structure (resembling Mali), so it’s not with out its dangers. We’ll simply have to attend and see the way it performs.

Talking of AI, I’ve crunched some numbers seen in Google’s inner paperwork and estimate that the DXT-45 is roughly 5% sooner at FMA operations than the G-715, which isn’t a complete lot. Nonetheless, it’s potential {that a} bigger 128-bit register means the DXT can nonetheless get extra carried out with every operation by way of SIMD and/or higher bandwidth utilization. It’ll be attention-grabbing to see if Google leverages the GPU for AI workloads, particularly as its TPU is simply taking a look at a 14% acquire subsequent era.

Nonetheless, I’m not satisfied that compute workloads or gaming efficiency are the rationale for switching — DXT doesn’t appear to be it’s going to beat the competitors right here. The true cause for the swap in all probability lies someplace within the stability between IP prices, vitality effectivity, and the function set on provide. Both approach, Google appears to have determined that Creativeness Applied sciences is the higher choice going ahead.

Sadly, those that hoped that the swap to a brand new GPU would propel the Tensor G5 and Pixel 10 up the graphics leaderboards might be dissatisfied. Whereas a modest 25% to doubtlessly a lot bigger 60% acquire may be very welcome, the chip will nonetheless lag two years behind the leaders. Worse, its GPU appears to be like set to stagnate once more with the Tensor G6, leaving the Pixel 11 even additional off the tempo.

Nonetheless, Tensor G5 is gaining a couple of instruments within the transition. Ray-tracing help, a unique structure for GPU-bound duties, and GPU virtualization imply that the Pixel 11 actually gained’t be brief on options and might be an improve for avid gamers. It would simply assist Google provide a couple of extra of these attention-grabbing Pixel-exclusive capabilities that hold the sequence within the highlight.

However that’s getting forward of ourselves; the Pixel 10 stays virtually a yr away and the competitors is already forging forward. Whereas Google is getting ready some CPU and GPU modifications with the Tensor G5, the chip remains to be, sadly, shaping as much as be a way behind the quickest processors within the enterprise. Will Google’s AI lead be sufficient to maintain its rivals at bay? I’m more and more involved that it gained’t.