MEDIATEK • DATA CENTER SOLUTIONS Data Center Solutions | AI Infrastructure MediaTek Data Center Solutions: Redefining AI Infrastructure Efficiency and Cost As the demand for AI is growing fast, data centers need to keep up. Hyperscalers running huge AI models are now building gigawatt-scale facilities. To support this, they need chips made specifically for AI, not general-purpose hardware. MediaTek brings its chip design expertise to data center solutions , offering a full path from custom chip design to rack-level deployment, while maximising efficiency and prioritising AI economics. Why Custom ASIC Design Matters Off-the-shelf chips are built to work for everyone. They are not built for one company's exact needs. This is why Custom ASIC Design is so valuable. Instead of forcing their AI workloads onto generic hardware, big cloud companies like Google, Meta, Amazon, and Microsoft are now building their own chips. These chips are designed around their specific models and power budgets. MEDIATEK • DATA CENTER SOLUTIONS Data Center Solutions | AI Infrastructure MediaTek supports this by working closely with chip foundries to fine-tune manufacturing processes. The company already has over ten chip designs made on the N3 process, a few more on N2, and its first A14 test chip is coming later this year. This experience comes from years of building mobile chips at massive scale, and MediaTek is now using that same know-how to build data center chips faster than many competitors. Building the XPU: From Wafer to Rack At the heart of MediaTek's approach is something called the XPU. This is a chip that is built specifically for AI work and is different from a regular CPU or GPU. MEDIATEK • DATA CENTER SOLUTIONS Data Center Solutions | AI Infrastructure Every XPU starts as a silicon wafer. This wafer is cut into smaller pieces called dies, each with its own job, like compute, memory, or input and output. These pieces are then joined together using advanced packaging. One method, called 2.5D packaging, places compute chips and memory chips side by side on a base layer. On the other hand, another method, called 3.5D packaging, stacks the pieces on top of each other. This shortens the distance data has to travel, which saves time and power. This matters a lot because memory speed, not just raw computing power, is often what slows AI systems down. MEDIATEK • DATA CENTER SOLUTIONS Data Center Solutions | AI Infrastructure MediaTek builds special memory interface chips that connect to high-bandwidth memory. The company is also exploring ways to process data directly inside the memory itself, which helps remove bottlenecks in large AI workloads. Networking Chips for AI: Keeping Everything Connected Fast chips are not enough on their own. They also need fast connections between them. This is where MediaTek's networking chips for AI come in. These chips connect components at every level. This includes chip-to-chip links inside a single package, board-to-board links inside a server, and rack-to-rack links across an entire data center. Technologies used include UCIe, high-speed SerDes, Ethernet, NVLink Fusion, and co- packaged optics. Co-packaged optics are especially important as AI clusters grow to include tens of thousands of chips. By placing optical components directly into the chip package instead of using separate cables, data centers can save power and reduce delays. This matters more than ever, since many facilities are now running into limits on how much electricity they can actually get. Real Partnerships, Real Results MediaTek's data center work is already happening in the real world, not just on paper. MEDIATEK • DATA CENTER SOLUTIONS Data Center Solutions | AI Infrastructure The company is a key partner in NVIDIA's NVLink Fusion ecosystem. This allows MediaTek's custom XPUs to connect directly with NVIDIA's platform, scaling out through InfiniBand and Ethernet networks. MediaTek has also teamed up with Microsoft to build a new type of Active Optical Cable using MicroLED technology. Early tests have shown strong power savings and longer reach for AI clusters, all while matching the reliability of traditional copper cables. Both companies are now working to bring this technology to mass production. Why Power Efficiency Is the New Priority for Data Center Chips MEDIATEK • DATA CENTER SOLUTIONS Data Center Solutions | AI Infrastructure For a long time, the main way to judge a chip was by how many operations it could do per second. That has changed. Now, the real question is how much useful AI output a chip can produce for each watt of power it uses. MediaTek treats power efficiency as a core part of its chip design, using lessons learned from years of building energy-efficient mobile chips. For companies planning their next AI data center, MediaTek's Data Center Solutions offer a full package. This includes custom ASIC chips , advanced packaging, and complete networking support, all designed to work together as one efficient system. FAQs What is MediaTek's Data Center Solution? It's a complete package of custom AI chips, packaging, and networking that MediaTek builds for large cloud and AI companies. What is Custom ASIC Design? It means building a chip from scratch to match one company's exact AI needs instead of using a generic, off-the-shelf chip. What is an XPU? An XPU is a chip built specifically to handle heavy AI tasks like training and running large AI models. How is an XPU different from a CPU or GPU? A CPU handles general tasks, and a GPU was originally made for graphics, while an XPU is built only for AI workloads.