Would future Pixels be better of returning to Snapdragon?


Unfortunately, internal projections from Google indicate that the Tensor G5 and Tensor G6 will follow a trajectory of little change.

Google’s subsequent two generations of smartphones are likely to leave power users dissatisfied in several aspects. Is it time for Google to revert to Snapdragon? Has the Tensor project already been deemed a failure? There appear to be increasingly compelling arguments in this regard.

Should Google continue to utilize Tensor processors?

1220 votes
Yes, Tensor is good. 33%
It should switch to Snapdragon. 35%
It should switch to MediaTek. 3%
Let’s give it a few more generations. 29%

A history (and future) of lagging behind
Robert Triggs / Android Authority
Fans of Pixel devices have grown accustomed to being distanced from cutting-edge performance metrics. Significant advancements in these areas did not emerge until the Tensor G3, and several key components, including the GPU and TPU, have remained unchanged with the newer Tensor G4. When you factor in subpar battery life and modem connectivity issues, particularly in earlier generations, Tensor's history is far from robust.

Further evidence of this can be observed in the canceled "redondo" iteration of the G4 — a more ambitious "fully custom" chip. Ultimately, this led to the "zumapro," which is merely a slightly enhanced variant of last year's "zuma" chip. Those anticipating substantial improvements following next year’s delayed shift from Samsung’s oversight may be disappointed. However, the chip will be more efficient, thanks to TSMC’s 3nm-class N3E manufacturing process.

The next two Tensor chips continue a trend of inconsistent performance enhancements. It is fair to note that recent Tensors are not sluggish for everyday tasks.

Nonetheless, with minimal performance upgrades across what will soon be three generations, it is challenging to identify where the additional capacity for exciting new features will originate.

In contrast, Qualcomm boasts a 45% peak NPU performance increase with its Snapdragon 8 Elite, along with increasingly tighter integration between the AI processing core and the image processor, enabling features like video object removal to be processed entirely on-device. While evaluating precisely how these competitors compare in terms of AI capabilities can be complex due to a variety of workloads, it is evident that upcoming phones powered by Qualcomm’s latest chip will possess a significant advantage in CPU and GPU tasks, potentially narrowing the gap or even surpassing Google’s machine learning capabilities. It likely won’t be long before features will follow suit. If this occurs, one must question the actual benefits of Tensor. This represents a substantial portion of a typical flagship's $999 MSRP and is likely impacting already narrow profit margins. Consequently, next-generation phones may see increased prices. Even Apple, which can produce processors for iPhones primarily at cost and benefits from substantial sales volume, is estimated to spend $100-$150 on its A18 Pro and the outsourced power and networking components.

Google cannot develop competitive chips without equaling the expenditure of its rivals.

Moreover, manufacturing at cutting-edge nodes, which is essential for achieving critical energy efficiency improvements, is becoming notoriously costly, especially as TSMC advances to next-generation 3nm nodes and beyond. Therefore, it will fall short in numerous performance metrics.

To address financial considerations, Google plans to reduce the silicon area of the Tensor G6 by approximately 8% compared to the G5. This will reportedly involve eliminating ray tracing from the GPU just a generation after its introduction, reducing a core from the DSP, and possibly eliminating the system-level cache, which is crucial for data exchange between the CPU and peripherals. The G6 is expected to introduce new, faster CPU cores; however, the architecture will narrow down to just seven cores, diminishing the upgrade's impact. In essence, Google is consciously choosing to downgrade or not enhance certain components of its chip to achieve cost savings and prioritize space for AI and imaging, as suggested by leaked plans. While all chips require compromise, Google’s limited budget means it faces more challenging decisions than its flagship competitors.

Google anticipates a modest CPU performance increase and approximately the same GPU performance as the preceding model, meaning the G6 will be regarded as another incremental upgrade. However, this situation may evolve, as the documents reviewed likely predate the final design approvals. Encouragingly, Google's documents indicate upcoming innovations in TPU and ISP; the company seems to be planning features such as 100x zoom and "Ultra Low Light video," along with enhancements in overall image quality and bokeh, as well as support for improved ML model compression. Furthermore, there is mention of a new NanoTPU for health and voice applications and a substantial overall TPU enhancement, expected to surpass the previous 14% advancement from the G3 and G4.

Given the focus on cost, the Tensor G6 seems to prioritize Google's dedication to AI and camera capabilities over raw performance. It remains to be seen if this strategy will suffice to keep competitors at bay, as they also work on their photography, AI, and overall performance upgrades in the coming years.

Google encounters typical issues in silicon development
Aamir Siddiqui / Android Authority
When reviewing Google’s Tensor initiatives comprehensively, it appears that the plans have suffered from the common pitfalls associated with custom silicon development — which can be highly expensive and presents numerous challenges in maintaining competitiveness. The abandonment of the "redondo" project appears to have influenced the project's overall trajectory, yet Google is not the first to experience scheduling delays.

Samsung faced prolonged challenges with its custom Arm-based Mongoose cores in the Exynos line, often falling short of achieving performance and efficiency objectives, ultimately abandoning the project in favor of Arm Cortex CPUs. Qualcomm also misjudged Apple’s rapid transition to 64-bit in 2013, ceasing custom CPU development of Arm's cores until the introduction of the 8 Elite’s Oryon CPU, which has since caused issues for its competitors. Custom development is inherently susceptible to missed deadlines and unexpected competitive advancements. Tensor appears to have encountered more obstacles than most in recent years, marked by inconsistent improvements and freezes in component development, rather than the steady yearly progress typically observed among competitors.

Missed deadlines and escalating costs are persistent challenges in silicon development. 

If Google is suddenly facing financial constraints related to silicon development, it suggests a degree of poor long-term strategic planning. The escalating costs associated with smaller manufacturing nodes have been known for years. Tensor encompasses a large, costly silicon area but has not leveraged this capacity to notably outpace its competitors. Additionally, licensing the most advanced components from Arm, Imagination Technologies, and others is not inexpensive (hence the mid-range GPUs), nor is the cost of developing TPUs and security chips independently.

Creating a high-caliber Tensor was always intended to be a substantial investment, yet the expenditures needed to be substantiated by tangible benefits. Benchmarks represent only a fragment of the overall assessment, and when evaluating real-world factors such as gaming performance and battery longevity, Tensor remains competitive and has shown notable advancements over time. However, it does not solely dominate the photography or AI markets. This might be somewhat harsh, considering the Pixel 9 is performing well in sales. Consumers generally prioritize performance specifications less, especially when applications and games function adequately (even if not at peak speeds), and battery performance is reasonably good now. Google effectively utilizes its custom TPU design to deliver impressive AI and camera functionalities for its Pixel series, which Google claims is a primary goal. The company intends to expand upon these features over the next few years, especially with the Tensor G6 and Pixel 11 launch approaching.

Nevertheless, Tensor may not be perceived as the success story that Google had anticipated and is increasingly viewed as a development burden rather than an asset. With challenges related to unreliable battery life, declining performance relative to competitors, and cost-reduction strategies, Tensor's past and future prospects appear uncertain at best.

Tensor is satisfactory at $799, but falls short of competitiveness at $999 and beyond.

The growing case against Tensor seems to underscore the fact that Google is charging higher prices than ever for the Pixel 9 series, while AI capabilities and point-and-shoot photography are merely part of the offering in the $1,000 price range. Even if this perception is unwarranted, power users and gamers are evidently underwhelmed with Google’s current offerings, and existing indications suggest that the Pixel 10 and 11 may not rectify this. Furthermore, what constitutes adequate performance today may not hold the same value in five, six, or seven years, especially as Tensor's inconsistent development cycle raises the risk of being overshadowed by rivals in upcoming iterations.

Nonetheless, Tensor’s distinctive features undoubtedly contribute to the appeal of entry-level Pixel flagships and the more affordable A-series, where performance demands are less critical. Rather than reverting to Snapdragon, it may be prudent for Google to explore a dual-tier chip strategy.