KernelBench-Hard analysis finds Grok 4.5 matches GPT-5.6 Sol on hardware roofline scores with superior paged attention performance · Digg