MediaTek's 2nm Dimensity 9600 Pro Pushes Flagship SoC Competition Toward Integrated Computing

The Dimensity 9600 Pro uses TSMC's 2nm process. Its CPU follows an all-big-core design, upgraded to a 2+3+3 layout: two C2-Ultra cores at up to 4.55 GHz, three C2-Pro cores at 4.35 GHz and three C2-Pro cores at 3.1 GHz. Total L2, L3 and SLC capacity reaches 34.5 MB, and the chip supports LPDDR6 and UFS 5.0. Compared with the previous generation, single-core performance rises 17% and multi-core performance rises 15%; at the previous generation's peak performance, multi-core power consumption falls 61%.

MediaTek executives framed the launch around agentic AI rather than raw speed alone. Xu Jingquan, a senior vice president at MediaTek, said at the event that “the computing baseline will no longer be the capability of a single core,” but will require integrating heterogeneous compute units into a system-level fusion compute architecture. Chen Yiqiang, deputy general manager of MediaTek's Wireless Communications Business Unit, said the industry used to ask how fast a flagship chip could run; in the agent era, it must also ask whether the chip can run that way all day.

The report said agentic AI changes how a phone works. Instead of waking the chip for short bursts when a user opens an app, takes a photo or starts a game, an agent needs to continuously perceive sound, images and the environment, understand requests in the background and sometimes execute tasks in advance. That can multiply daily wakeups and make CPU idle periods shorter and more fragmented. Power may be consumed less by one heavy calculation than by repeated core startups, task switching and data movement. AI also cannot monopolize the phone: while a user games, watches video or joins a meeting, background agents may still run, and large models can compete with apps for memory while CPU and NPU hand off tasks.

MediaTek director, president and COO Chen Guanzhou said AI entry points will expand from PCs to phones and other terminals, and compute will move from centralized cloud deployment to a hybrid of cloud and edge. AI users will expand from enterprises to consumers. A phone is not only an entry point to cloud AI factories; it can also become a distributed AI factory that produces tokens, he said. Privacy is a direct constraint. Chen said “some private data cannot be sent to the cloud” and can only produce tokens in the terminal AI factory.

To address those constraints, MediaTek placed CPU-NPU collaboration, memory and flash data exchange, and scheduling of compute, power and memory resources into one design. The Dimensity 9600 Pro's all-big-core power scheduling is measured at the microsecond level, and its fast core on-off capability is 10 times that of the previous generation. System-level intelligent compute power management aims to reduce background losses that users do not easily notice. When AI and traditional apps run together, memory becomes a bottleneck. MediaTek uses compression and dynamic scheduling to improve memory efficiency; according to MediaTek, large-model startup speed increases 27% while 25 apps can remain alive.

The chip keeps a dual-NPU design. MediaTek's second-generation ultra-energy-efficient NPU integrates a low-power sensor hub and mainly runs lightweight models, with power consumption down 40% from the previous generation. It handles continuous understanding of sound, motion and the environment for proactive agents. The ultra-performance NPU 1090 handles heavier model computation, with large language model prefill performance up 51% and tokens generated per watt up to 55% higher. The new generation 3.0 generative AI engine doubles INT4 computing capability and supports Hybrid Attention, new model compression technology and on-device operation of MoE models up to 30B parameters. MediaTek says the 30B scale allows on-device models to call more complex expert capabilities while MoE activates only part of the parameters, reducing compute and memory pressure.

MediaTek pointed to applications with vivo and OPPO. A vivo camera assistant can use the low-power NPU to continuously understand the viewfinder and proactively recommend features and composition. An OPPO LUMO inspiration composition feature analyzes the positional relationship between subject and environment in real time. An agent can also reorganize information scattered across hotels, flights, schedules and to-do lists and proactively remind users before they ask. Chen Yiqiang summarized MediaTek's AI approach as three elements: always-on operation, sufficient intelligence and an application ecosystem. The chip supports the latest Gemini Nano, and MediaTek is working with Google on an on-device AI privacy protection architecture that isolates, tracks and centrally manages agent access to personal data.

AI also crosses into GPU, ISP and communications. The G2-Ultra NX GPU has 27% higher peak performance than the previous generation, consumes 24% less power at the previous generation's peak performance and improves ray tracing performance by 18%. MediaTek describes the key upgrade as a neural rendering architecture in which the GPU, AI compute units and memory data paths participate in rendering to improve image quality within mobile power budgets. After nearly a year of joint development with the Delta Force project team, MediaTek brought mobile ray tracing from fixed scenes to a complex dynamic battlefield. CAT, technical director of the Delta Force project team, said quality improvements should not come at the expense of frame rate; reflections in scopes, dynamic ground reflections and lighting on different materials must run stably in a real game before they become player experience.

Imaging follows the same pattern. The Dimensity 9600 Pro integrates the Imagiq 1290 ISP and builds an intelligent imaging architecture through ISP-NPU fusion. The NPU handles multi-frame composition, semantic analysis and subject recognition, while the ISP handles zone-based color and tone processing. Full-depth computational photography, 60fps sports-level tracking, 17EV high dynamic range, 4K60 cinematic direct output, 4K120 Log color video and 4K240 slow motion are results of that coordinated architecture in different scenarios.

AI is also used in cellular networks, Wi-Fi and positioning. The chip runs network models based on the real-time environment and dynamically selects base stations and communication parameters. In complex cellular environments, stall duration falls 30% from the previous generation, and can fall a further 40% when combined with an OEM's large model, according to MediaTek. Positioning accuracy improves 30% over the previous generation. AI thus moves from a generative feature users see directly to a foundation hidden in networks and systems.

The Dimensity 9600M launched alongside the Pro. It carries the AI agent engine, scheduling engine, smart imaging and neural rendering capabilities, and is meant to bring this system design to more product tiers. The report said the significance of the 9600 series is not only that MediaTek reached 2nm first or added another product. Data center AI competition has already moved from buying more accelerator cards to coordinating compute, storage, interconnect and software. A similar shift is happening on phones, with constraints changing from rack power and networking to a few watts, one battery and limited memory. As agents continuously sense, reason and execute, stronger CPU, GPU and NPU units are only an entry ticket; how they work together will determine the experience. That also changes how flagship chips are judged. Traditional benchmarks can still measure the ceiling of a compute unit, but they struggle to reflect the real cost of agents waking repeatedly through a day, running alongside multiple apps and continuously calling personal data. More persuasive metrics may include task completion time, tokens generated per watt, app retention, and temperature and battery life during long runs. MediaTek's 2nm launch is therefore a starting point. The larger signal is that flagship mobile SoCs are moving from a single-core performance race into a competition over system efficiency, where the strongest chip must not only run fast but also know when to compute, who should compute, and how to keep computing all day.