AI HW SW CoDesign Workstream
- Project Lead
- Weifeng Zhang
Rapidly evolving artificial intelligence (AI) continues to drive ever-increasing computation demands. However, innovation in AI models and the capability to train and infer from these models are often constrained by compute platform capabilities. Since system hardware design cycles are long and design goals may be simplified to reduce development cycles, optimization of these systems for performance and operational efficiency will be imperative.
Today’s systems do not offer high levels of performance and flexibility while meeting performance and operational efficiency targets. Thus, this AI co-design workgroup aims to innovate and enable a scalable yet composable computing infrastructure which can dynamically adapt to the application's runtime characteristics and execute the application with high efficiency.
In 2023, a vision addressing the current system shortcoming was outlined in a whitepaper titled “Polymorphic Architecture for Future AI Applications.” It proposed using hardware-software co-design principles with some key requirements, including large-scope scalability and polymorphic computing infrastructure with hierarchical transformability and composability.
Development of these ideas continued through deeper discussions of architecture, system design considerations, evolution of the State of the Art in AI models and performance expectations. The result of OCP's Community effort culminates into the new whitepaper which describes, in greater detail, a Polymorphic Architecture for AI computing that unleashes unprecedented flexibility, efficiency, and scalability.
Polymorphic Architecture represents a paradigm shift in how we conceptualize and implement AI computing systems. At its core are two key principles:
- Continuous transformability
- Recursive composability
These principles enable computing resources—from individual chiplets to entire data centers—to dynamically reconfigure themselves, optimizing performance for diverse AI workloads in real-time.
Scope
Using the Polymorphic Architecture as an exploration platform, AI co-design workgroup continues its study in the following technology areas:
- Seamless scaling across computing hierarchies, from chips, systems, to large-scale clusters
- HW/SW co-design to improve resource (e.g. compute and interconnect) utilization and energy efficiency
- Flexible configuration and adaptation of resources for future-proofing against rapidly evolving AI algorithms
- Means to realistically utilize heterogeneous resources in systems
- Potential for significant cost savings in AI infrastructure
As outlined in the white papers above, foundational concepts, technical specifications, and potential applications of Polymorphic Architecture will continue to be part of focus to accelerate AI computation. The group also has a high-level roadmap for implementation, including simulation frameworks and APIs for standardization. As the AI landscape continues to evolve, Polymorphic Architecture offers a forward-thinking solution to the challenges of next-generation AI computing. Industry leaders, researchers, and developers are invited to explore this transformative approach and contribute to shaping the future of AI infrastructure.
Get Involved
There are several ways to get involved in the OCP Future Technologies Initiative:
- Join the mailing list to receive updates: FTI Mailing list
- Join the mailing list of this workstream: AI HW-SW CoDesign Mailing list
- Join the conversation in the FTI LinkedIn Group
- Connect with the FTI team and the community, give advice and contribute.
- Join the OCP Future Technologies Symposium, submit a paper, propose participants, attend!
OCP FTI - Hardware Software CoDesign Calendar
The calendar displayed here is updated nightly from the project's Groups.io Calendar