AMD engineers have cracked the code for multi-chiplet gaming graphics cards after years of struggling with latency problems. The chip giant filed a patent revealing breakthrough technology that could revolutionize how graphics processors handle data across multiple computing units. Their secret weapon involves a smart switching system that connects different chiplets without the delays that plagued previous designs. The patent details show AMD plans to use scaled-down versions of their successful Infinity Fabric technology from data centers. Gaming enthusiasts may finally get the performance boost they've been waiting for.
The company's new approach centers around a data-fabric circuit that makes split-second decisions about where to send graphics tasks. Each graphics compute die gets its dedicated L1 and L2 cache memory, while a shared L3 cache connects all the chiplets together. The smart switch compares requests in nanoseconds to determine whether tasks need migration or data copying. AMD borrowed ideas from their AI accelerator success stories to make consumer graphics cards more powerful. The shared staging zone between chiplets mimics their popular 3D V-Cache technology from processors.
AMD already has the manufacturing partnerships and software ecosystem ready for mass production. TSMC's advanced packaging technology will help build these complex multi-chiplet designs at scale. The company plans to merge gaming and AI architectures under their upcoming UDNA platform. Previous RDNA 3 graphics cards suffered from chiplet communication delays that frustrated gamers. The UDNA 5 generation might be the first to deliver AMD's chiplet gaming vision to consumers.
The company's new approach centers around a data-fabric circuit that makes split-second decisions about where to send graphics tasks. Each graphics compute die gets its dedicated L1 and L2 cache memory, while a shared L3 cache connects all the chiplets together. The smart switch compares requests in nanoseconds to determine whether tasks need migration or data copying. AMD borrowed ideas from their AI accelerator success stories to make consumer graphics cards more powerful. The shared staging zone between chiplets mimics their popular 3D V-Cache technology from processors.
AMD already has the manufacturing partnerships and software ecosystem ready for mass production. TSMC's advanced packaging technology will help build these complex multi-chiplet designs at scale. The company plans to merge gaming and AI architectures under their upcoming UDNA platform. Previous RDNA 3 graphics cards suffered from chiplet communication delays that frustrated gamers. The UDNA 5 generation might be the first to deliver AMD's chiplet gaming vision to consumers.
