MediaTek’s Dimensity 9400 processor promises 28% greater multithread performance than the previous-gen flagship while employing a similar computing subsystem. At the same time, the smartphone processor raises peak GPU performance by 40%. The biggest gains, however, come in AI processing. MediaTek claims large language models (LLMs) are 80% faster than before and that it’s ushering in an agentic-AI era, the follow-on to the gen-AI era.
The Prime-Core Directive
The D9400’s computing subsystem comprises eight big CPUs like last year’s Dimensity 9300. However, some big CPUs are bigger than others. An Arm Cortex-X925 replaces the largest, fastest prime Cortex-X4 in the D9300. Clocking 11% faster and delivering 15% greater per-cycle throughput, the X925—in conjunction with a bigger Level 2 cache and a speedier DRAM interface—raises Geekbench single-thread performance by 35%. Like the D9300, the new smartphone processor also integrates three additional X4s and four Arm Cortex-A720s. However, the D9400 raises their clock rate and increases their L2 cache to boost performance.
The Cortex-A720 is a curious choice. The Arm Cortex-A725 isn’t much better, but it is newer. Moreover, adopting the TSMC N3E process, MediaTek had to retarget the A720—why not start with the fresher A725? We suspect that the company expends significant effort optimizing its licensed cores, and much of this work would carry forward to the new process. Because the A725 changes so little, it made more sense to reuse the A720, which could meet the company’s performance and power targets. The Cortex-X4 also hearkens back to Arm’s previous generation. In addition to the factors affecting A720 selection, MediaTek may have chosen it because it’s smaller than the X925.
LoRA and LoRa
The D9400 integrates MediaTek’s eighth-gen NPU 890 to accelerate AI functions. New capabilities include on-device low-rank adaptation (LoRA), multimodal AI, and the ability to run an agentic engine. LoRA enhances a neural network without retraining the full model, enabling customization with a small additional data set. Multimodal AI refers to a model that accepts input from multiple data sources, such as text and images. Behind the LoRA and multimodal capabilities lie NPU-architecture tweaks and software development. Software also accounts for the Dimensity Agentic AI Engine (DAE), which provides a framework and standard APIs to on-device models.
The GPU upgrades the D9300’s 12-core Immortalis-G720 to the new Arm Immortalis-G925. Improved ray tracing speed and various new features such as opacity micromaps (OMM) continue to narrow the gap between smartphone and PC graphics. Camera and multimedia functions also improve, and the D9400 has three MIPI ports for trifold displays such as those found in accordion-fold phones.
Lest we forget, for end users looking to make calls, the D9400 updates communications functions. The 5G modem conforms to 3GPP Release 17. A companion connectivity chip reduces Wi-Fi and Bluetooth power and increases the radios’ ranges. MediaTek claims phone-to-phone direct Bluetooth connections are possible up to 1.5 km, which is as baffling as it is impressive.
Bottom Line
The Dimensity 9400 is MediaTek’s third flagship smartphone processor, and the company has rapidly grown the business. It expects 50% growth this fiscal year over the $1 billion level set in 2023. Customers include Chinese brands Vivo, Oppo, Xiaomi, and iQoo, which are drawn to MediaTek’s simple value proposition: top-tier features and performance at a lower price than Qualcomm.
In addition to taking share, MediaTek has the same challenge as others in the smartphone industry: how to motivate consumers to upgrade their devices. The common hope is that AI-based features will excite the market. The image-generation and LLM demos of past years are interesting parlor tricks but nothing more. Nebulous as it is, the agent concept could solidify into a useful product. Although AI hype borders on the comical, we can’t fault MediaTek for seeking to advance the technology to a new era.