Frequently Asked Questions

ClankerBuilder is a PC configurator specifically designed for AI and LLM workstations. It combines a compatibility engine with real-world benchmarks to help you build systems that can actually run your target AI models. Unlike generic PC builders, we validate VRAM requirements, check PCIe lane allocation, ensure adequate power delivery, and provide tok/s (tokens per second) performance estimates.

Our tok/s ratings come from community benchmarks and lab tests, never our own measurements. Each estimate includes confidence levels and source attribution. We aggregate multiple data points when available and clearly mark when estimates are extrapolated. The goal is transparency - you see exactly where each number comes from and how confident we are in it.

It depends on your target models. Smaller models like 7B parameter LLMs can run on consumer GPUs with 8-12GB VRAM. Larger models like 70B parameters typically need multiple high-end GPUs or expensive single cards with 40GB+ VRAM. Our compatibility engine helps you match hardware to your specific model requirements.

GPU VRAM capacity is typically the primary constraint. Most AI models need to fit entirely in GPU memory for good performance. After VRAM, consider memory bandwidth, then compute performance. CPU, RAM, and storage are important but usually not the bottleneck for inference workloads.

Yes, but with caveats. For inference, you can split large models across multiple GPUs, but you need NVLink or high-bandwidth PCIe connections for good performance. For training/fine-tuning, multiple GPUs are very beneficial. Our compatibility engine checks PCIe lane allocation and power requirements for multi-GPU setups.

Used GPUs can offer excellent value for AI workloads. Cards like the RTX 3090 (24GB VRAM) or Tesla V100 still perform well for many models. We include used/refurbished pricing in our recommendations when available. Just verify the card's condition and ensure your PSU can handle the power requirements.

For most AI inference workloads, 32-64GB of system RAM is sufficient. The GPU handles the model, so system RAM mainly stores data preprocessing and system operations. For training larger models or running multiple models simultaneously, you may need 128GB or more.

VRAM capacity is #1 - if your model doesn't fit, performance crashes. After that: memory bandwidth (crucial for autoregressive generation), compute units (for parallel processing), and memory hierarchy (L2 cache, HBM vs GDDR). CPU and system RAM matter less for pure inference.

Currently, NVIDIA has better software ecosystem support (CUDA, cuDNN, better framework compatibility). AMD GPUs can work but may require more setup and have fewer optimizations. For serious AI work, NVIDIA is usually the safer choice, especially for newer models and frameworks.

AI workloads often run GPUs at 100% utilization for extended periods. Ensure adequate case airflow, consider aftermarket GPU coolers for high-end cards, and monitor temperatures. Thermal throttling will hurt performance more than almost any other factor.

Use our cloud vs. buy calculator to compare costs for your usage patterns. Generally: heavy daily use (>4 hours/day) favors ownership, light or experimental use favors cloud. Factor in electricity, cooling, and your time value. Cloud gives you access to cutting-edge hardware without upfront investment.

It varies by budget and requirements, but RTX 4070 Ti SUPER (16GB) and RTX 4090 (24GB) often provide good VRAM/$ ratios for new cards. For used markets, RTX 3090 and A6000 can offer excellent value. Our value ratings account for both performance and current pricing.

AI hardware moves quickly. Focus on builds that handle your current models well rather than over-engineering for unknown future needs. VRAM requirements grow over time, so prioritize generous VRAM over raw compute if choosing between them. Plan for 2-3 year refresh cycles for cutting-edge work.

Our compatibility engine flags potential issues: insufficient PSU wattage, PCIe slot conflicts, RAM limitations, or cooling concerns. Red flags mean the build likely won't work. Amber warnings suggest monitoring or potential limitations. Green means good compatibility across all checked parameters.

Check the model's VRAM requirements against your GPU(s). For multi-GPU setups, ensure your interconnect bandwidth supports the model size. Our model pages show hardware compatibility for popular LLMs, including minimum and recommended specs.

First, verify all components match the exact specifications. Check power connections, ensure adequate cooling, and update GPU drivers. Our compatibility engine is conservative, but edge cases exist. Check our methodology page for details on how compatibility is determined.