{"slug": "gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed", "title": "Gemini 4 Neon wrote AMD kernels that nearly doubled Inkling 2 Mini decode speed in 4.3 hours", "summary": "An AMD engineer used Gemini 4 Neon to port Thinking Machines Lab's Inkling 2 Mini to a Strix Halo machine in about 4.3 hours, producing custom Loom kernels that ran decode at 44.8 tokens per second versus 23.8 tokens per second for llama.cpp on ROCm, and reached 5.9x llama.cpp prefill at an 18k-token prompt. The agent-built Loom package was measured against what the AMD engineer called \"super nascent\" day-one support for the hardware.", "body_md": "# Gemini 4 Neon wrote AMD kernels that nearly doubled Inkling 2 Mini decode speed in 4.3 hours\n\n**The agent-built Loom package also hit 5.9x llama.cpp prefill at an 18k-token prompt, against day-one support the AMD engineer called super nascent.**\n\n        By [Ryan Merket](https://runtimewire.com/author/ryan-merket)\n        · Published \n\nAn AMD engineer used Gemini 4 Neon to port Thinking Machines Lab's Inkling 2 Mini to a Strix Halo machine in about 4.3 hours, producing custom Loom kernels that ran decode at 44.8 tokens per second against 23.8 tokens per second for llama.cpp on ROCm.…", "url": "https://wpnews.pro/news/gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed", "canonical_source": "https://runtimewire.com/article/gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed", "published_at": "2026-11-22 01:50:37+00:00", "updated_at": "2026-11-22 02:14:05.415084+00:00", "lang": "en", "topics": ["ai-agents", "large-language-models", "ai-infrastructure", "ai-tools", "machine-learning"], "entities": ["Gemini 4 Neon", "AMD", "Thinking Machines Lab", "Inkling 2 Mini", "Strix Halo", "Loom", "llama.cpp", "ROCm"], "also_reported_by": [], "takeaway": "Agent-written Loom kernels lifted Inkling 2 Mini decode on Strix Halo from 23.8 to 44.8 tokens per second over llama.cpp on ROCm and hit 5.9x llama.cpp prefill at an 18k-token prompt, in about 4.3 hours of agent work.", "recommendation": {"advice": "On new AMD hardware with nascent ROCm support, have a coding agent generate custom Loom kernels rather than relying on llama.cpp defaults; the source measured 44.8 vs 23.8 tokens/s decode and 5.9x prefill at 18k tokens.", "playbook": "inference-serving", "playbook_url": "https://wpnews.pro/playbooks/inference-serving"}, "alternates": {"html": "https://wpnews.pro/news/gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed", "markdown": "https://wpnews.pro/news/gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed.md", "text": "https://wpnews.pro/news/gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed.txt", "jsonld": "https://wpnews.pro/news/gemini-4-neon-wrote-amd-kernels-that-nearly-doubled-inkling-2-mini-decode-speed.jsonld"}}