MediaTek’s Dimensity 9600 Pro combines two Arm C2-Ultra and six C2-Pro CPU cores with the Mali G2-Ultra NX GPU. This hardware enables agentic AI to scale from cloud systems to physical robotics and mobile devices. (agentic AI: artificial intelligence capable of acting and coordinating across cloud, edge, and physical environments.) It matters to system architects who need to maintain power and thermal limits in flagship smartphones.
Is this for you?
Robotics and mobile hardware engineers should adopt Arm's unified compute foundation to manage power, thermal limits, and real-time coordination for agentic AI.
- If you are a robotics hardware designer, use the unified foundation to manage complex system workloads and real-time control tasks.
- If you are a mobile SoC engineer, note that the Dimensity 9600 Pro powers vivo X500 Pro and OPPO Find X10 devices.
- If you are a game engine developer, implement NFRU to reduce GPU rendering load and improve motion.
- Skip this if you are a general smartphone consumer looking for software-only AI updates.
- Skip this if you are a software-only developer seeking high-level API documentation.
What you can do now
- If you are a mobile SoC engineer, compare Mali G2-Ultra NX performance against the Adreno GPU for neural graphics.
- If you are a chiplet system architect, verify that the Arm compute platform interconnects meet the PCIe Gen 6 requirements.
- If you are a chiplet system architect, monitor the tape-out date to ensure the chiplet design is finalized.
Key figures
- Dimensity 9600 Pro CPU Configuration
- 2 Arm C2-Ultra and 6 C2-Pro cores
- SME2 Performance Gain
- 60% higher performance for toxic-speech detection
- Audio Processing Speed
- 15 seconds of audio processed in 0.37 seconds
- Optimization Challenge Submissions
- 330+ submissions
Arm’s Connected Computing Continuum for Agentic AI
Arm’s platform creates a “connected computing continuum” that unifies cloud, edge, and physical AI into a single foundation. This shift allows diverse workloads, models, and software environments to coordinate across different environments rather than operating as separate systems. The strategy provides a consistent compute foundation that helps developers move intelligence from where it is created in the cloud to where it is acted upon in the physical world.
Robotics hardware designers
Robotics hardware designers, who coordinate sensing, planning, and safety in real-time, should use this unified foundation to manage complex system workloads. A consistent platform helps ensure that data movement and real-time control tasks are handled efficiently enough to maintain safety in real-world deployments. This holds as long as the platform can maintain predictable performance across all three layers of the continuum.
MediaTek Dimensity 9600 Pro Core Configuration
MediaTek’s Dimensity 9600 Pro integrates two Arm C2-Ultra and six C2-Pro CPU cores with the Mali G2-Ultra NX GPU. These flagship-tier components provide the foundation for AI, graphics, and responsive computing in high-end mobile devices. The configuration is designed to help manufacturers stay within the power and thermal limits that define modern smartphone design.
Mobile SoC engineers
Mobile SoC engineers who configure the GPU to utilize integrated neural accelerators should note that this chip powers devices like the vivo X500 Pro Series and OPPO Find X10 Series. (SoC: a single integrated circuit that contains all the necessary components of a computer or smartphone.) These specific launches demonstrate how Arm CPU and GPU advances move from design into rich device experiences. If the hardware fails to maintain these thermal limits during sustained AI workloads, the design fails to meet flagship requirements.
Mali G2-Ultra NX and Neural Frame Rate Upscaling
Mali G2-Ultra NX GPU architecture includes dedicated neural accelerators sitting directly within the GPU silicon. These accelerators enable techniques like Neural Frame Rate Upscaling (NFRU), which uses AI-generated intermediate frames to reduce rendering demands and improve motion. By moving this processing onto the GPU, developers gain more headroom to balance visual quality and power consumption.
Game engine developers seeking to improve mobile gaming frame rates
Game engine developers seeking to improve mobile gaming frame rates should implement NFRU to reduce the rendering load on the GPU. The integration of neural hardware directly into the GPU allows for real-time graphics improvements without exceeding power and thermal limits. This holds while the software tools for NFRU remain compatible with existing rendering pipelines; a move to a proprietary, incompatible rendering path would break the integration.
AMBA C2C and UCIe Chiplet Interoperability
AMBA C2C provides the chip-to-chip connectivity for chiplet systems by separating protocol and packetization from the underlying die-to-die transport. (chiplet: an integrated circuit component that is part of a larger multi-chip system.) (AMBA C2C: a chip-to-chip connectivity standard for chiplets that separates protocol and packetization from transport.) This separation allows chiplet designs to remain scalable and interoperable across different vendors and generations. Because it supports the UCIe standard, AMBA C2C allows for the composition of multi-chip systems without forcing every design into a fully coherent architecture.
Chiplet system architects
Chiplet system architects who manage multi-chip system interoperability should configure the AMBA C2C interconnect to link processor and memory chiplets. Using this standard allows the design to remain interoperable with other vendors using UCIe. The risk of being wrong is a non-scalable design that requires a full coherent architecture, increasing complexity and cost unnecessarily.
SME2 and 60% Performance Gains for Voice Safety
Arm SME2 is an instruction set extension that provides a 60% performance boost for toxic-speech detection on a single CPU thread. (SME2: an instruction set extension used to optimize specific workloads like toxic-speech detection.) By running this workload on a single thread, the system leaves compute headroom for other applications, such as gaming. The combination of SME2 with LiteRT, quantization, and KleidiAI allows 15 seconds of audio to be processed in 0.37 seconds on a test device. (quantization: the process of reducing the precision of a model to run more efficiently on hardware.)
Edge AI developers
Edge AI developers who need to implement real-time safety for a social app should use the combination of LiteRT, quantization, KleidiAI, and Arm SME2. This approach allows for faster on-device moderation with less reliance on server processing while preserving security. If the required model cannot be successfully quantized for these specific instructions, the performance gains will not be realized.
Physical AI and the Arm AGI CPU
Arm AGI CPU provides the compute foundation needed to coordinate sensing, inference, planning, and control for physical AI. Federico Pecora, Senior Principal Robotics Research Lead at Arm, notes that the challenge for robots is making these distinct workloads work together reliably in the real world. The goal is to translate greater intelligence into predictable, efficient, and safe operation.
Robotics systems engineers designing robots for warehouse navigation or hospital
Robotics systems engineers designing robots for warehouse navigation or hospital use should configure the robot’s control system to run on the Arm AGI CPU. This helps ensure that the data movement for LiDAR sensor data and other sensing inputs is handled by a foundation built for complex coordination. If the sensing-to-planning latency exceeds real-time limits, the robot’s operation may become unsafe.
The Zara Robotics Startup Case Study (Worked Example)
Zara runs a robotics startup building a robot that must navigate a crowded hospital and detect patient needs in real-time. The team chooses the Arm compute foundation to coordinate sensing, planning, and control, helping ensure that the robot can process LiDAR sensor data and perform inference on the Arm AGI CPU. By using this integrated foundation, the robot achieves predictable and safe operation in a complex environment.
The situation flips if the startup chooses a non-integrated system where sensing and planning run on separate, uncoordinated processors. In that case, the complexity of coordinating these distinct workloads might lead to latency issues or safety risks, potentially causing the robot to fail its real-time requirements.
Replacing Server-Side Workloads
On-device AI and the Arm compute platform reduce the need for server-side voice moderation and manual rendering of high-frame-rate graphics. By moving toxic-speech detection to the device via SME2, manufacturers reduce reliance on server processing while preserving security. Neural Frame Rate Upscaling (NFRU) reduces the need for manual rendering of every frame, reducing the GPU load while improving motion via AI-generated frames.
Who this is not for: General smartphone consumers looking for software-only AI updates
General smartphone consumers looking for software-only AI updates should skip this technical analysis and instead check product reviews for the vivo X500 Pro or OPPO Find X10. Software-only developers seeking high-level API documentation should instead consult the Android or Linux developer documentation.
Arm compute platform: what to check, compare and watch
- Adreno GPU
- The Arm compute platform's Mali GPU is a primary alternative to the Adreno GPU in mobile SoC design. Mobile SoC engineers: compare the Mali G2-Ultra NX performance against the Adreno GPU for neural graphics tasks.
- LiteRT performance metric
- LiteRT performance is a key metric for running on-device models on the Arm compute platform. Mobile SoC engineers: watch the LiteRT performance metric to ensure it meets the required latency for on-device speech processing.
- NVIDIA Jetson module
- The Arm compute platform is a primary alternative to the NVIDIA Jetson module for robotics hardware. Robotics hardware designers: compare the Arm compute platform cost against the NVIDIA Jetson module for the robot's sensing unit.
- PCIe Gen 6
- The Arm compute platform's interconnects must often align with external standards like PCIe Gen 6. Chiplet system architects: verify that the Arm compute platform's interconnects meet the PCIe Gen 6 requirements.
- Tape-out date
- The tape-out date is the deadline for finalizing the chiplet design on the Arm compute platform. Chiplet system architects: monitor the tape-out date to ensure the chiplet design is finalized on the Arm compute platform.
Arm compute platform: questions the original report leaves open
- What is the central claim, and who is making it?
- The original report did not say what the central claim is.
- What evidence supports it?
- The original report did not say what evidence supports the claim.
- What is the strongest argument against it?
- The original report did not say what the strongest argument against it is.
- What would prove the claim wrong?
- The original report did not say what would prove the claim wrong.
- What are the specifications of the Arm AGI CPU?
- The original report did not say the specifications of the Arm AGI CPU.
- What is the bandwidth of AMBA C2C?
- The original report does not say.
Key terms
- Agentic AI
- AI that can act and coordinate across cloud, edge, and physical environments.
- Neural Frame Rate Upscaling (NFRU)
- A technique using AI-generated intermediate frames to reduce rendering demands.
- AMBA C2C
- A chip-to-chip connectivity standard for chiplets that separates protocol and packetization.
- SME2
- An Arm instruction set extension used to optimize specific workloads like toxic-speech detection.
Original reporting: Top Arm-based innovations advancing AI infrastructure, on-device intelligence and physical AI in September 2026 (Arm), on 5 October 2026.