GPU and Mac prices for local LLMs

We check public product pages one at a time in an ordinary browser, the way a shopper would, only at shops whose robots.txt allows it, and record the price, the page URL and the date. We don't run scrapers against shops. Shops used so far: Apple, NVIDIA Marketplace, Newegg, momo購物網, 原價屋 CoolPC, 欣亞數位 Sinya.

Each price is what the product page listed that day, for a new, in-stock single card or computer: the cheapest listing we found at that shop, and at most two shops per device per region. Taiwan prices are as listed, which includes business tax; US prices are before sales tax and shipping.

The latest price is the lowest price on the most recent check day. NT$ and US$ are separate markets, and we never convert between them.

Value = the estimated generation speed for Llama 3.1 8B Instruct (Q4_K_M, 8K) from our estimator, divided by the price in units of NT$1,000 or US$100. It measures how much memory bandwidth the money buys; it isn't a benchmark, and it ignores power, noise and software support. For Macs and mini PCs the price is for the configuration named, and the price per GB uses that configuration's memory.

Next: official shop APIs (eBay for used prices, Best Buy) once API keys are approved, plus weekly manual checks imported from a CSV file. The data is free to reuse under CC BY 4.0: download the JSON or use the API, and please link back.

Methodology →