AMD Instinct MI400: New Details on Architecture and Multi-Chiplet Design | Hardware computer list | How much cpu and gpu do i need | Cpu-z | Turtles AI
AMD is preparing the launch of Instinct MI400 accelerators, which will mark a significant evolution in design with an innovative architecture based on multiple chiplets and separate active interposers. The new solutions will aim for superior performance in AI, with improved efficiency and optimized communication between processing units.
Key points:
- Advanced architecture: The Instinct MI400 will adopt CDNA Next, with possible developments toward the UDNA designation.
- Innovative design: There will be up to eight XCDs with more efficient interconnections via active interposers.
- Communication evolution: Implementation of distinct I/O and multimedia tiles for optimized data flow management.
- Progress over MI350: Expected improvements in AI performance and efficiency through a more advanced production node.
AMD is preparing to innovate the AI accelerator industry with the next generation of Instinct MI400, introducing a fundamentally revamped architecture designed for greater efficiency and unprecedented performance. Although the company is currently focused on the launch of the MI350, scheduled for 2024 and based on the CDNA 4 architecture, significant details are already emerging regarding the future MI400, expected in 2026. From the latest patches spotted on Free Desktop, key information emerges that outlines an important structural evolution. The new accelerator will feature up to eight XCDs (Accelerated Compute Dies) spread across two active interposers (AIDs), effectively doubling the number compared to the MI300 design. Unlike previous generations, which relied on the Infinity Fabric for inter-die communication, the new configuration also features dedicated I/O and multimedia tiles designed to optimize data transmission between compute units and memory and connectivity interfaces. The architecture on which the MI400 will be based is referred to as CDNA Next, but inside sources suggest that it may be renamed UDNA as part of a strategy to unify RDNA and CDNA technologies, strengthening cohesion between high-performance computing solutions and the gaming world. In parallel, the closer release of the Instinct MI350 will already mark a major step forward, with a 3nm manufacturing process that will provide greater energy efficiency and up to a 35-fold increase in AI inference performance over the previous model.
Although full technical details of the MI400 have not yet been made official by AMD, the direction taken by the company suggests an acceleration in the development of solutions for large-scale AI computing, with an increasing focus on efficiency and scalability.