https://www.msuiche.com/categories/software-engineering/Recent content in Software Engineering on Matt SuicheHugoen-usSun, 28 Sep 2025 00:00:00 +0000https://www.msuiche.com/posts/multi-gpu-programming-with-amds-iris-framework-for-triton/Sun, 28 Sep 2025 00:00:00 +0000https://www.msuiche.com/posts/multi-gpu-programming-with-amds-iris-framework-for-triton/<p>GPU production constraints are creating infrastructure bottlenecks. Multi-GPU programming, particularly vendor-agnostic implementations, has become essential. In their <a href="https://www.youtube.com/watch?v=H2bzSn5ZPks" target="_blank" rel="noopener">GPU Mode presentation</a>, AMD Research engineers Muhammad Awad, Muhammad Osama, and Brandon Potter introduced Iris—a Python library that enables fine-grained multi-GPU programming in Triton. Similarly to my previous <a href="https://www.msuiche.com/posts/gluon-when-triton-isnt-low-level-enough/">Gluon blogpost</a>, this post captures my understanding and interpretation of their work, serving as both technical documentation and personal reference for this emerging multi-GPU programming paradigm.</p> <h2 id="technical-problem">Technical Problem <a href="#technical-problem" class="anchor">🔗</a></h2><p>Current multi-GPU programming uses bulk synchronous models (BSP) through libraries like NCCL. This model enforces sequential phases:</p>