Team builds megakernel outperforming llama.cpp on RTX 5090

The author argues that the current AI boom is driven less by hardware improvements and more by mathematical and algorithmic optimization, achieved by instructing AI models to intentionally make mistakes and explore alternative paths. Testing a custom megakernel created through this method on an RTX 5090 GPU yielded significantly higher token-per-second performance than standard frameworks like llama.cpp. The author concludes that software optimization enables developers to extract unused capacity from existing hardware rather than relying solely on expensive hardware upgrades. Open article →