[Question] Roadmap, TileOps Integration and TileLang Collaboration

[PengYoun9]

Hi TileFoundry maintainers,

Thanks for open-sourcing TileFoundry. After reading through the architecture and specification documents, I have a few questions about the project's longer-term roadmap and its relationship with other projects in the Tile-AI ecosystem.

1. TileFoundry roadmap

Is there a public roadmap for TileFoundry?

Such as: Support for additional hardware instructions and GPU architectures

2. Integration with TileOps

Will high-performance kernels generated by TileFoundry eventually be integrated into TileOps?

For example, is the expected workflow something like:

• TileFoundry generates and evaluates candidate kernels

• Selected kernels are validated and tuned

• Production-ready implementations are integrated into TileOps

Or are TileFoundry and TileOps intended to remain independent, with TileFoundry primarily serving as a compiler and kernel-generation research platform?

3. Relationship with TileLang

Is there any planned collaboration or integration between TileFoundry and TileLang?

The two projects appear to address related but different layers:

TileLang provides a user-facing GPU kernel programming and compilation environment.
TileFoundry introduces its own HIR/TIR, shard/layout abstractions, target code generation, and an architecture that appears well suited to agent-driven kernel generation.

Will you use TileFoundry to generate or optimize TileLang programs?

More generally, how do the maintainers see the respective roles of TileFoundry, TileOps, and TileLang within the Tile-AI ecosystem?

Thanks for any clarification.

[zhen8838]

Hi,

Thanks for your detailed questions and your interest in TileFoundry and the broader Tile-AI ecosystem. We’re glad to clarify these points for you:

1. TileFoundry is currently an experimental project focused on exploring how agents and compilers can collaboratively optimize tensor programs. As our first step, we are carrying out initial build and validation work on NVIDIA GPUs.
2. Yes, that is the plan. The original vision for TileFoundry was to build an agent-driven automatic optimization pipeline for TileOps.
3. TileFoundry will work in tandem with TileLang. Specifically, it will leverage TileLang’s existing compilation and optimization capabilities to reduce the complexity of agent tuning, and draw on TileLang’s mature multi-backend support to enable multi-device compatibility. The operators produced through this workflow will ultimately expand and enrich the TileLang ecosystem.
4. Built on TileLang as its foundational layer, TileFoundry serves as a platform for agent-driven automatic kernel optimization. The final optimized kernels will be contributed and stored in TileOps.

We will share more updates as the project moves forward. Please feel free to reach out if you have any further questions or feedback.

Best regards,
The TileFoundry Maintainers

[PengYoun9]

Thank you for the detailed clarification. My understanding is now:

• TileLang: provides the foundational compilation, optimization, and multi-backend capabilities
• TileFoundry: provides the agent-driven automatic kernel optimization
• TileOps: stores and distributes the final validated and optimized kernels

This direction sounds very promising, especially because reusing TileLang’s compiler and multi-backend infrastructure should allow TileFoundry to focus more directly on the agent search, optimization, validation, and feedback loop.

It will also be interesting to see how the current TileFoundry HIR/TIR and specification-driven architecture evolve as the integration with TileLang becomes deeper—for example, whether they will remain as an agent-facing planning and optimization IR, or gradually be lowered into or aligned with TileLang’s existing compiler representations.

Thanks again for sharing the vision. I look forward to following the project’s progress and future updates.