Hardware Sizing Calculators & Guides for AI Infrastructure
Xingyu Zhitong is a small set of calculators and guides for people who buy, build or run hardware for AI work. The calculators answer practical sizing questions: how much GPU memory a model needs once the KV cache is included, how big a power supply a workstation needs, what a GPU server costs to power, when buying beats renting cloud GPUs, and how long a battery will run a Jetson robot. Every calculator runs in your browser, shows the formula it uses, and links to a guide that explains the assumptions. The figures come from vendor datasheets and model configuration files, and each page says where the estimate stops being reliable.
Start here
Deploying local LLMs and sizing VRAM
Calculate model parameter footprints and dynamic attention KV cache memory across sequence lengths and quantization levels.
Workstation and server power planning
Determine continuous wattage, transient excursion headroom, and ongoing utility electricity charges for GPU systems.
Infrastructure procurement and cloud economics
Assess capital payback horizons, colocation hosting fees, and breakeven utilization thresholds when buying versus renting GPUs.
Edge computing and robotics runtime
Estimate battery discharge duration and flight endurance for Jetson Orin modules operating under active deep learning workloads.
All 19 guides by topic
Guide index →Will it fit? Sizing GPU memory
Weights, KV cache, quantization and context length: how to work out memory before you buy a GPU, with measured numbers from our own T4 run.
- How to Estimate GPU Memory for an LLM: Weights, KV Cache and Overhead
- KV Cache Memory Explained: Why Long Chats Eat Your VRAM
- FP8, INT8 and 4-bit Quantization: What You Save and What You Risk
- What a 128K Context Window Really Costs in Memory and Compute
- Which GPU Do You Need for a 7B or 8B Language Model?
- Fine-Tuning Memory Requirements: Full Training vs LoRA vs QLoRA
Specific models and machines
Worked hardware plans for named models, from a 70B model at home to Llama 405B and DeepSeek-V3, plus what Apple's unified memory can hold.
Speed and multi-GPU scaling
Why decoding is limited by memory bandwidth, how to split a model across GPUs, and which inference server fits which job.
Power, cooling and running cost
Power supplies and connectors, data-center PUE, heat removal and the break-even point between renting and owning GPUs.
Data-center GPUs and edge devices
What actually changes between A100, H100, H200 and B200, and how long a battery runs a Jetson AGX Orin.
How we work out the numbers
Our calculators are built on published hardware datasheets, transformer architecture specifications, and empirical power delivery standards. Memory equations calculate parameter bit widths and attention key-value caches directly from model layer dimensions. Power estimates follow ATX 3.0 excursion specifications and data center PUE metrics rather than marketing ratings. We operate independently without hardware vendor sponsorships or affiliate bias. To review our complete calculation methodologies and references, read our editorial policy or learn more about our mission.