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

We’re barely via processing the arrival of the Pixel 9 collection, 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 a much 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 may not be a reputation you’re tremendous aware of in at present’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 keep 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 improvement.

To be blunt, Google’s Tensor collection underwhelms within the graphics division, languishing at the very least two generations behind the quickest within the enterprise when it comes to efficiency. It’s additionally been gradual to undertake new GPU designs and continues to dodge ray-tracing assist, a distinct segment characteristic however one which we now count on in a flagship-tier cellular GPU. That appears set to alter, at the very least 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 with regards to 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 pace. 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, at the very least on this check. That might be essentially the most important leap within the Pixel’s graphic efficiency for generations, however we’d count on an excellent bigger bounce 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 present and upcoming 2025 rivals packing the powerhouse Snapdragon 8 Elite.

Google’s inner figures, as seen by Android Authority, recommend efficiency might bounce a bit greater. The graphs aren’t nicely labeled however level to a 35% to 60% achieve 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 may nonetheless make for an fascinating comparability level. Fortunately, the DXT structure sports activities some fascinating options that may shut the hole on its rivals.

Ray tracing stays elective with DXT, as it’s with Arm’s Mali/Immortalis cut up. Google is choosing the smallest Ray Acceleration Cluster (RAC) unit configuration it could actually (a DXT-48-1536-0.5RT2), with a half RAC in every core. Once more, the G5 is just not 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 (liberating 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 mentioned, 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 fascinating tidbits. The structure helps as much as 2×4 and 4×4 Fragment Shading Price (aka Variable Price 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 assist. The important thing takeaway is that it is a GPU structure that’s very aggressive from a characteristic standpoint.

We additionally know the brand new GPU helps virtualization, which isn’t present in present Tensor chips. This enables for the usage of accelerated graphics in a digital machine, doubtlessly permitting Google to convey one among 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 fascinating 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 models 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 variety 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 commonest graphics knowledge sizes, with smaller knowledge sizes optionally supported inside these ALUs for machine studying. The normal setup is nice for graphics and never unhealthy for decrease bit-depth machine studying workloads both. Nonetheless, it could actually’t leverage single instruction a number of knowledge (SIMD) on bigger knowledge varieties, which may be useful for reminiscence bandwidth and cache assets, that are all the time at a premium in cellular GPUs.

Paired with two GPU cores that work independently additionally means doubtlessly greater efficiency and/or decrease energy consumption when crunching via graphics and compute workloads, due to parallel processing efficiencies. In different phrases, you possibly can have cores contribute to a single or completely different workload as quick as doable or energy down a core to avoid wasting vitality.

Further effectivity financial savings for graphics and/or machine studying workloads might have caught Google’s eye. That mentioned, these cores can’t share inner assets, which might result in bottlenecks or underutilization in comparison with a unified shader structure (comparable to 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 doable {that a} bigger 128-bit register means the DXT can nonetheless get extra finished with every operation via SIMD and/or higher bandwidth utilization. It’ll be fascinating to see if Google leverages the GPU for AI workloads, particularly as its TPU is just taking a look at a 14% achieve subsequent era.

Nonetheless, I’m not satisfied that compute workloads or gaming efficiency are the explanation for switching — DXT doesn’t appear like it’s going to beat the competitors right here. The true purpose for the swap most likely lies someplace within the steadiness between IP prices, vitality effectivity, and the characteristic set on supply. Both method, Google appears to have determined that Creativeness Applied sciences is the higher possibility 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% achieve 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 number of instruments within the transition. Ray-tracing assist, 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 players. It’d simply assist Google supply a number of extra of these fascinating Pixel-exclusive capabilities that preserve the collection within the highlight.

However that’s getting forward of ourselves; the Pixel 10 stays virtually a 12 months away and the competitors is already forging forward. Whereas Google is getting ready some CPU and GPU adjustments 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 opponents at bay? I’m more and more involved that it gained’t.