Leading the Generative AI
Technology Revolution
Generative AI is driving the next wave of innovation in semiconductors.
As the industry leader in developing high performance and power-efficient system-on-chip products, MediaTek is already bringing the benefits of our powerful, in-house designed AI processors into our diverse product ecosystem.
We now ship over 2 billion chips in connected devices every year, many of which contain our AI processors, making us one of the world’s leading edge AI suppliers. With our broad market reach, you’ll find our AI technologies in devices at home, in education, in businesses, in mobile, and in vehicles.
MediaTek Hybrid AI – Edge and Cloud
MediaTek is uniquely positioned to drive and support the adoption of Generative AI applications and technologies, whether deployed on the cloud, at the edge in devices, or as a hybrid approach that takes advantage of both the cloud and the edge together.
The Comprehensive MediaTek AI Ecosystem
The MediaTek AI ecosystem includes hardware, development tools, and a software development kit (SDK).
Software developers have access to comprehensive tools and algorithms, while MediaTek NeuroPilot allows a write-once-deploy-everywhere strategy that simplifies app development across MediaTek’s entire product range.
The MediaTek NPU: AI Processing Unit
The MediaTek NPU, or processor dedicated for AI tasks, is as important as the CPU and GPU within our system-on-chips.
The MediaTek NPU provides a very efficient architecture to implement AI acceleration in edge devices, such as smartphones, tablets, AI+IoT, smart TVs, networking SoCs, and more.
Using five common neural networks as a benchmark, the average power/performance curves show that MediaTek's Deep Learning Accelerator (DLA) contained with the NPU offers 27x greater power efficiency compared with a typical CPU and 15x greater power efficiency compared with a typical GPU.
In 2023, MediaTek launched its 7th generation NPU, which is specifically designed to accelerate generative AI based on transformer models. Its highly flexible design can scale performance with computation units, power use, memory bandwidth, and memory capacity depending on the application requirements, whether it's in a smartphone SoC or part of a data center-grade superchip.
What makes a MediaTek NPU?
The MediaTek NPU is a highly scalable multi-core AI processor that can contain both MDLA and MVPU cores in varying quantities, depending on the application requirements.
Each MediaTek NPU contains the following:
- Hardware-based multicore scheduler.
- A dedicated DMA engine that performs deep layer fusion and data compression to reduce the demand on DRAM bandwidth.
- Power intelligence and shared memory awareness.
- MediaTek designed Network on Chip (NoC) for low latency intercore communication.
MediaTek Deep Learning Accelerator (MDLA)
- High performance, programmable and versatile, with an energy-efficient MAC architecture that can perform data reuse in a wide-range of NN applications.
- Designed to handle different types of networks, including:
- (a) Convolutional Neural Networks
- (b) Recurrent Neural Networks
- (c) Long Short-Term Memory models and
- (d) mobile Bidirectional Encoder Representations from Transformers (BERT)
- Supports INT4/INT8/INT16, FP16, BF16 and mixed precision to further reduce power consumption.
- Inter subsystem direction communication with the ISP and DPU (display) minimizes latency and power consumption.
MediaTek Vision Processing Unit (MVPU)
- General purpose DSP optimized for computer vision (CV) and neural network (NN) applications.
- Best balance of power and performance for visual processing applications such as photography, videography, and video stream playback.
MediaTek NeuroPilot SDK
MediaTek NeuroPilot is a set of software tools and APIs for customers and developers to optimize their AI-based applications for use on MediaTek SoCs.
Both local and cloud versions of these tools are available, including pre-trained deep learning models that can further accelerate development. NeuroPilot supports many mainstream AI frameworks including TensorFlow, Caffe, ONNX, and transformers, and it works with both Linux and Android OS. During the post-silicon stage, developers can deploy their trained AI model on their platform to measure the performance directly.
Efficiency at its core
NeuroPilot also offers platform aware optimization. With NeuroPilot an application will understand the SoC hardware it operates on, sending the AI-tasks to the most appropriate processing unit, whether it's the CPU, GPU, NPU, or DSP.
In addition, where necessary, NeuroPilot will even adopt heterogeneous computing to combine the strengths of multiple processing units to provide the best results — for example, using a GPU and NPU together in 3D graphics rendering AI-VRS, AI-HDR, or super resolution upscaling.
MediaTek Research is a specialized AI research team that includes accomplished researchers with diverse backgrounds in computer science, engineering, mathematics, and physics. Our vision is to push the limits of what is possible in Artificial Intelligence (AI) and Machine Learning (ML), which, we believe, can unlock new opportunities, discoveries, and progress that will shape our future.