بچه‌های دیتابریکس یه سیستم باحال به اسم «پروتئوس» راه انداختن که کارش تولید خودکار کرنل‌های اختصاصی برای پردازنده‌های گرافیکیه

https://www.databricks.com/blog/achieving-extreme-efficiency-through-specialized-gpu-kernel-generation We recently published a blog on work on Proteus, a system for generating specialized GPU kernels. Why should models from 1B to 1T parameters rely on the same generic kernels? By specializing for the exact shapes encountered at runtime, we can unlock extreme efficiency. Using Proteus, we generated individual kernels for Qwen 3.5 122B that were 1.8–5.2× faster than the best implementations available in vLLM. But finding a fast kernel is only the beginning. The harder problem is knowing whether its results can be trusted. Generation is often the cheap step in terms of token cost. In our initial harness, knowledge and context management, not candidate generation, consumed most of the token budget. Verification introduces another cost: candidates must run on real GPUs, in isolation, often more than once. Verification must be airtight. A kernel must be functionally correct, but its performance claims also need rigorous validation: fair comparisons, multiple timers, clean state, hidden tests, and repeated measurements. Otherwise, we generate impressive numbers and not production-ready kernels. Agentic kernel generation makes extreme specialization possible. Building a reliable, production-ready system around it remains a hard frontier. At @Databricks, we’re tackling these challenges at the intersection of AI and systems to make inference faster, more efficient, and autonomously improving. @leshenj15 Previous AA results from Databricks posted by @Yuchenj_UW : https://x.com/Yuchenj_UW/status/2094993931268420072?s=20