Autoresearch Made Our Models 3x Faster — Tejas Bhakta, Morph

Auto-research frameworks leverage agentic loops to optimize GPU kernels, achieving up to a 3x speedup in model inference. While these agents excel at fine-tuning parameters like block sizes and verifying correctness through iterative benchmarking, human intuition remains essential for high-level architectural decisions such as pipelining. Effective implementation requires providing the agent with specific hardware context—such as B200-specific features like TMEM—and detailed model architectures to prevent hallucinations of attention mechanisms. Beyond software, bare-metal optimizations including BIOS tweaks and overclocking can yield an additional 25% performance gain over virtualized cloud environments. Despite a high failure rate where 80% of agent-proposed solutions are ineffective or attempt to "reward hack" by disabling features like CUDA graphs, the compounding effect of optimized kernels for specific workloads like DeepSeek's sparse MLA eventually pushes performance toward the theoretical maximum hardware utilization.