Learn How the Flagship Mobile Chips com

After a previous version of this guide put face-to-face chips from Qualcomm, MediaTek, Apple, Samsung, and Huawei/HiSilicon, a bunch of things changed in the smartphone component arena, notably the arrival of Google, the death and rebirth of flagship chips from Huawei due to US sanctions, and Samsung going on a hiatus from the flagship space. First, let’s take a look at an overall specs comparison of today’s contestants.

Snapdragon vs. Dimensity vs. Exynos vs. Tensor vs. Apple: Specs Compared

SoC parts

Mobile system-on-chip (SoC) design has been pretty stable in the past few years, with most chip manufacturers following a similar multi-core design with many specialized blocks dedicated to offloading processing tasks from the central processing unit (CPU):

• Graphics Processing Unit (GPU): Responsible for accelerating visual calculations, especially 3D rendering.
• Image Signal Processor (ISP): Process data captured from the camera sensors.
• Digital Signal Processor (DSP): Process data obtained from other sensors, sometimes employed for media and/or communication processing.
• Neural Processing Unit (NPU) (aka AI Accelerator, APU for MediaTek): Specialized in artificial intelligence and machine learning tasks.
• Modem: Responsible for adding support for cellular networks (e.g. 4G/LTE, 5G Sub-6 GHz, 5G mmWave…).

CPU cores are commonly split between efficiency cores and performance cores, with the first consuming considerably less energy but not particularly fast. Heavier tasks are offloaded when necessary to the performance cores, with Android chips adopting a third tier of even faster CPUs denominated as “Prime cores”.

For almost 10 years, all smartphones from traditional brands have used the same CPU architecture: ARM. However, not all ARM CPUs are equal. Apple, for example, designs its own CPU cores compatible with the ARM instruction set, while the Android suppliers adopt reference CPU cores from ARM, known as the Cortex cores.

With the success of Apple’s in-house designs, both Qualcomm and Samsung are expected to once again offer custom ARM designs, something that mirrors the early days of Android SoCs with Qualcomm Krait and Samsung Mongoose cores. As of 2023, ARM’s reference cores for mobile processors are split between Cortex-X, Cortex A7xx, and Cortex-A5xx cores for prime, performance, and efficiency cores, respectively.

In the past three years, Android flagship SoCs had one main differentiating factor when compared to mid-range chips: The prime cores. Qualcomm’s launch of the Snapdragon 7+ Gen 2, however, made that distinction obsolete, but that chip still features a slightly less capable GPU, ISP, and modem compared to the company’s flagship offerings.

Production process

Apple had the first consumer processors made on TSMC’s 7-nm—technically announced after a Huawei Kirin chip, but reaching stores first—, 5-nm, and 3-nm, the latter used for the current A17 processor used on the iPhone 15 Pro models.

MediaTek also has a strong partnership with TSMC, but usually is one process node behind Apple. In the past, the company was known for not having energy-efficient processors, but that was influenced by using outdated fabrication processes from both TSMC and UMC, also from Taiwan. The Dimensity chips, on the other hand, are fabbed using modern processes, making them competitive in energy consumption.

Qualcomm, for its part, usually maintains a foundry-agnostic strategy, with production booked according to availability. The Snapdragon 888/888+ and 8 Gen 1, for example, were made by Samsung Foundry, but their successor, the Snapdragon 8+ Gen 1 uses TSMC’s N4 process, resulting in less power consumption and heat dissipation.

Samsung’s in-house chips are obviously done by Samsung Foundry, which for the past few years has not been competitive with TSMC’s nodes from the same “nm” generation. The South Koreans are optimistic, however, that its 3 nm-class process will offer a 45% improvement in power usage, which would be a welcome evolution in mobile phones.

As for Google, since the Tensor mobile chips are based on Samsung LSI’s products, its processors are also fabbed by Samsung Foundry, with rumors suggesting a future switch to TSMC, time will tell. Tensor chips usually distinguish themselves from Exynos SoCs by employing different CPU and GPU clusters, and most importantly, Google’s in-house Tensor NPU.

GPU trends

The first one was AMD joining the mobile SoC space with its Radeon GPUs on the Samsung Exynos 2200, using the same RDNA2 architecture—although with far fewer cores—on the latest PlayStation and Xbox consoles. Samsung’s retreat from the flagship mobile SoC space, however, meant that it didn’t have a follow-up until two years later, with the Exynos 2400 chip used on the Galaxy S24 models packing an RDNA3 GPU.

Qualcomm’s Adreno GPUs follow their own path, with very few publicly available details when it comes to specifications. Recent generations support ray tracing, alongside other advanced features, making them also suitable for use in other Snapdragon chips designed for Windows laptops.

MediaTek mostly uses ARM’s reference GPUs in its Dimensity range, with the newer Dimensity 9200/9200+ powered by the RT-capable Immortalis-G715 GPU. In the past, the Helio SoC family featured considerably fewer ARM Mali GPU cores than its Exynos and Kirin rivals, leading to lower performance, something that the flagship Dimensity models are changing thanks to a more premium positioning and advanced TSMC nodes.   

Google doesn’t quite fit with other SoC brands in the Android space, going on a different pace on GPUs just like on the CPU cores. The Tensor G2 features a relatively low seven GPU cores from an older generation, so it is incompatible with the same features available on its rivals, with the same recipe used on the G3 processor from 2023.

Apple doesn’t disclose much about its in-house GPU designs, which share a lot of history with old-school PowerVR video processors. iPhone GPUs are nevertheless very capable in terms of features, sharing their core design with the larger Apple M chips used on Mac PCs.

AI everywhere

With artificial intelligence (AI) and machine learning (ML) even more in the spotlight than in previous years thanks to Stable Diffusion and ChatGPT, the importance of AI-dedicated cores is only expected to increase. In that regard, flagship mobile chips have been getting increasingly more powerful in ML tasks thanks to dedicated cores such as Apple’s Neural Engine, Qualcomm’s Hexagon, MediaTek’s APU, and Google’s TPU.

Unfortunately, manufacturer-advertised numbers are not comparable between brands, with Apple announcing 16 trillion operations per second (TOPS) on its A16 Bionic chip, while Qualcomm’s numbers used to include processing done between the DSP, CPU, and GPU until the company stopped disclosing TOPS numbers for its Snapdragon chips.

Regardless, expect companies to keep advertising AI features on chips, throwing bold claims on performance improvements, while operating systems and apps try to catch up in supporting them.

Diminishing returns?