Hey Product Hunt 👋 I kept getting burned the same way: rent a GPU to try a model, watch it OOM or crawl at half the speed the vendor page implied, and eat the bill anyway. Vendor spec sheets don't tell you if a specific model + quant + context length actually fits and runs fast on a specific card, so most people just guess, or rent-and-pray. GPU Sizer is the tool I wished existed: deterministic VRAM math, not a rough rule of thumb, plus a throughput model, validated to an 11.3% median error against 19 published third-party benchmarks, out-of-sample, every source linked so you can check it yourself. Pick a model (83 tracked, or paste any HuggingFace repo), a quant, a context length and concurrency, and it ranks 100+ GPUs by best fit, best value, and max performance, with VRAM, cost, power, and TTFT for every row. It also goes past "does it fit": Designer lets you co-locate multiple models on a GPU pool, and Factory turns your setup into a P&L — cost per million tokens, break-even vs. the API you'd otherwise pay, margin at your real utilization. Free forever, no card required. Would genuinely love your feedback, especially if you break it. What's the workload you want to size first?

GPU Sizer

Know which GPU runs your LLM, validated on real hardware

Know which GPU runs your LLM, validated on real hardware