New project: GPUQuicklist.com

Here I am belatedly announcing a new project called GPUQuicklist.com. It is a site that lets the user quickly see the live prices for graphics cards they select. I created it about 6 months ago, but I thought I’d get around to writing up the motivation for creating it.

Along with everyone else it seems, the release of OpenAIs chatgpt really got me interested in LLMs. Then came LLama from Meta (Facebook). When Georgi Gerganov introduced llama.cpp, it became possible to run the llama models on your own machine, provided you could fit them into memory. Much of the early excitement around llama.cpp was around the speed at which it could run llama models on Apple Silicon.

At a quite early stage of llama.cpp, Quantisation and CUDA support were added: 2023-03-30 for 4bit quantisation and 2023-04-03 for GPU support, which is where my interest in getting a good price for a graphics card comes in. If you have a PC already, a graphics card is cheaper than an Apple system upgraded to 32GB of RAM for example. (I suppose it may possible to run 4-bit quantised llama 8b on a 16GB Mac, but I wanted to upgrade to 32GB of RAM to run certain databases)).

When I found out about diskprices.com I enjoyed the speed at which the user could find relevant information about harddrives on amazon. If you have searched for products on amazon, you’ll know that many products returned will only tangentially resemble what you asked for. This occurs because Amazon’s algorithm relies on a combination of keyword matching, user behavior, and product popularity. It doesn’t use all keywords from your search to match some items, but includes popular items that are only loosely related to increase their chances of a sale. With diskprices.com, you can look at a specific product type and be fairly sure that you are seeing the cheapest prices available on amazon.

Diskprices.com lets the user save the products they’re interested in by bookmarking the page. The user’s selections in the checkboxes are saved to the url, which makes the bookmarking functionality available. This is pretty neat, and means we don’t have to worry getting users to login, nor any cookies! Of course I provide the same feature for gpuquicklist.com.

GPU Quicklist is my creation, inspired by diskprices. Like diskprices for Harddrives, gpuquicklist.com allows you to see a compact table of GPU prices ordered by price. It has a bunch of filters by GPU model and also can filter by GPU RAM. I have included current and last generation Nvidia and AMD graphics cards as well as the current Intel Arc GPUs. I also have a pro gpu page, which is where you can find professional level cards and a few server cards.

I really enjoyed being able to see a list of the current best prices and being able to filter rapidly and see the results instantly on a very fast site. Diskprices.com seems to use a perl script for getting the data by country, whereas I have created a static page for each country, simply because I prefer generating static pages to using a perl server page. Either seem to be able to make use of caching via a CDN for speed, which is great user experience.

The data has been fetched via the Amazon Advertising API, which allows product data to be searched. Currently I have access to the API for USA, UK, Germany and Spain. I was briefly able to get some Australia product data a month or two ago, but the data stopped updating almost a month ago, unfortunately. I still have the listings for AU at gpuquicklist.com/au.

One problem with using the API is that there are occasionally keyword matches that creep into the product listings that aren’t graphics cards. This mainly occurs for water blocks or replacement fans. The listing would say cooler for such and such a card, for example:

90mm T129215SU Video Card Cooler Fans with Case Compatible for ASUS Gaming
 GeForce RTX 3070 3060TI V2 8GB GDDR6 ATS-RTX3060TI Graphics Card 

Currently I manually remove some of these listings, but I should really get around to introducing some “smarts” into the process, such as using gpt-40-mini to check, or to have an exclusion phrase list, such as “compatible for” which would detect the product above.

So there it is, I have introduced the website gpuquicklist.com. I enjoyed making it, but there is still some work to do to clean up spammy product listings.