<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Milind Gunjan — Writing</title><link>https://gunjanlabs.com/writing/</link><description>LLM inference, telecom and AI infrastructure.</description><item><title>GPU utilization lies</title><link>https://gunjanlabs.com/writing/gpu-utilization-lies/</link><guid>https://gunjanlabs.com/writing/gpu-utilization-lies/</guid><pubDate>Sun, 22 Nov 2026 12:00:00 -0700</pubDate><description>Your dashboard says 100%. Your users say the chatbot takes four seconds to answer. Both are right.</description></item><item><title>Prefill vs. decode</title><link>https://gunjanlabs.com/writing/prefill-vs-decode/</link><guid>https://gunjanlabs.com/writing/prefill-vs-decode/</guid><pubDate>Sun, 22 Nov 2026 12:00:00 -0700</pubDate><description>Every LLM request has two phases, and they are bottlenecked by completely different things.</description></item><item><title>Continuous batching</title><link>https://gunjanlabs.com/writing/continuous-batching/</link><guid>https://gunjanlabs.com/writing/continuous-batching/</guid><pubDate>Sun, 22 Nov 2026 12:00:00 -0700</pubDate><description>Decode reads the weights once per step, no matter how many users share it. That is the good news hidden in the bottleneck.</description></item><item><title>KV cache memory sets concurrency</title><link>https://gunjanlabs.com/writing/kv-cache-memory/</link><guid>https://gunjanlabs.com/writing/kv-cache-memory/</guid><pubDate>Sun, 22 Nov 2026 12:00:00 -0700</pubDate><description>If batching is where throughput comes from, memory decides how big the batch can get.</description></item><item><title>Putting it together</title><link>https://gunjanlabs.com/writing/putting-it-together/</link><guid>https://gunjanlabs.com/writing/putting-it-together/</guid><pubDate>Sun, 22 Nov 2026 12:00:00 -0700</pubDate><description>Four ideas, one chain. How they fit on one GPU, what each optimization pulls on, and what to actually watch.</description></item></channel></rss>