How RigFit calculates

VRAM needed

Text models: weights + KV cache + 1.0 GB overhead. Weights = parameter count × bits per weight (Q4 ≈ 4.85 bits like llama.cpp Q4_K_M, Q8 ≈ 8.5, FP16 = 16). KV cache = 2 × layers × KV heads × head dim × context × 2 bytes (FP16 cache), with sliding-window layers capped at their window. Parameter counts come from Hugging Face safetensors metadata, and architecture from each model's config.json; both are cached in data/models.json.

Image/video models: diffusion weights at the chosen precision + VAE/activation allowance for the stated resolution; text encoders assumed offloaded to system RAM.

Tokens per second (estimate)

Generating each token reads all active weights plus the KV cache from VRAM, so speed ≈ 0.6 × memory bandwidth ÷ bytes per token. It's a ballpark for single-user chat, not a benchmark.

Buy prices

Updated daily by scripts/update_prices.py from public price trackers. Every price on the site shows its source and date. Where no current price was found we show the launch MSRP, labelled as such, and don't rank the card. Latest run:

• bytesizeddeals: ok (8 prices)
• videocardprices: ok (30 prices)
• gpusold: ok (26 prices)
• siliconcomps: ok (23 prices)
• amazon_paapi: disabled (no API keys)
• ebay_browse: disabled (no API keys)
• policy: dropped 12 Amazon-attributed prices (no Creators API access yet)

Sources: ByteSizedDeals (new, RTX 50 / RX 9000), VideoCardPrices (new retailer + lowest used listing), GPU Sold (median used eBay sold price incl. shipping; Silicon Comps asking-price medians as fallback). GPU specs are hand-curated from vendor spec sheets (cross-check: TechPowerUp GPU Database).

Rental prices

Updated daily by scripts/update_rentals.py: Vast.ai public marketplace search (cheapest verified, rentable, on-demand single-GPU offer with ≥95% reliability), Runpod's published pricing page (schema.org offers, Community and Secure Cloud), and Lambda's published on-demand table (1x GPU instance). Latest run:

• vastai: ok (23 prices)
• runpod: ok (18 prices)
• lambda: ok (5 prices)

Break-even hours = cheapest buy price ÷ cheapest hourly rental. It ignores electricity, resale value and the rest of the computer; the calculator on the Rent page lets you add those.

Builds

GPU and CPU prices only (CPU: ByteSizedDeals CPU Price Checker). Other parts are listed but not priced until we have a legitimate source. Daily price history is kept in data/history/; buy/wait signals need 14 days.