On Tuesday, at the Computex 2024 event held in Taiwan’s Taipei, Mediatek shared its vision on the future of AI computing. The company plans to put generative AI-capable chips on smartphones, automobiles, and servers. The company also announced the development of AI applications in the data centre, infrastructure systems, telecommunications and more based on Arm Neoverse CSS (Compute Subsystems). Rick Tsai, the vice chairman and CEO of Mediatek, stated, "AI starts at your hand, on your mobile device/smartphone, and we want AI to be around you, and we want to be a major player in enabling this." "A phone (vivo X100S) that you carry in your pocket can deliver the AI performance of the latest Copilot+ PC," the CEO added. Tsai also shared that the company's flagship MediaTek Dimensity 9300+ offers up to 68 trillions of operations per second (TOPs) of combined AI performance (GPU+NPU). The neural processing unit (NPU) itself can offer up to 48 TOPs of AI performance, which makes it faster than Snapdragon X Elite, and as powerful as the Intel's upcoming Lunar Lake laptop processor. It is also compliant with the requirement of Copilot+ PCs, which requires a processor with an NPU with at least 40 TOPs of AI performance. Dimensity 9300+ also enables on-device AI experiences like an AI eraser. The company also demonstrated the world's first on-device LoRA Fusion demonstration on a smartphone which can generate animations in real-time along with on-device dynamic image generation SDLX turbo, which can generate images and preview it in real-time. Tsai also confirmed that Mediatek Dimensity 9400 SoC will be announced in the next three months. The upcoming Mediatek Dimensity Auto Cockpit C-X1, designed in collaboration with NVIDIA TensorRT and CUDA RTX GPU, with support for DLSS and ray-tracing, is showcased to be capable of delivering up to 5.8 times more AI performance (TOPs) with 4.5 times faster LLM (large language model) performance, along with 5.1 times less latency. MediaTek's upcoming AI accelerator meant for edge computing will be fabricated using TSMC's 2nm architecture, and it is said to have over 70 billion transistors. In comparison, the 4nm Dimensity 9300+ has over 22.7 billion transistors. The writer is in Taiwan, Taipei, on the Invite of Intel India.
