Credo and AMD Break AI Memory Bottleneck with Programmable Logic Accelerators
Event summary
- Credo showcased high-performance memory scaling for programmable logic-based AI accelerators at the 2026 OCP Global Summit.
- The demonstration featured OmniConnect Weaver memory fanout gearbox on an AMD Versal AI Edge Series Gen 2 adaptive SoC evaluation board.
- The system enables terabytes-class LPDDR memory with high bandwidth and capacity over VSR links without requiring silicon interposers.
- Application software runs unmodified as the expanded memory is transparent above the controller.
The big picture
As AI inference workloads grow, memory bandwidth and capacity have become critical bottlenecks, often limiting performance more than compute throughput. Credo's OmniConnect Weaver addresses this by enabling scalable memory architectures on programmable logic platforms, potentially disrupting the dominance of custom ASICs in high-performance AI applications. This collaboration with AMD highlights the industry's shift toward more flexible, high-bandwidth memory solutions for next-generation AI systems.
What we're watching
- Adoption Pace
- How quickly AI developers will integrate Credo's memory solution into their programmable logic designs.
- Performance Gains
- Whether the memory scaling will translate into measurable improvements in AI inference workloads.
- Market Differentiation
- The extent to which this collaboration will position Credo and AMD against custom ASIC solutions.
Related topics
