Start with what you’re trying to do
All three tools
Local LLM GPU Comparison
Ranks GPUs for running language models locally by what matters in that order: whether the model fits in VRAM, then tokens per second per dollar, power draw and setup effort.
PC Parts Value Comparator
Scores graphics cards by usable value for a chosen job — 1080p, 1440p or 4K gaming, local AI, workstation, budget or low-power builds — instead of one raw benchmark number.
GPU Mining ROI Calculator
Daily, monthly and yearly mining profit for a GPU after electricity and pool fees, with the break-even time and the coin that pays best for that card right now.
For local AI, memory comes before speed
Gaming benchmarks rank cards by frames per second, and that ranking is the wrong one for running a language model on your own machine. The first question is not how fast a card is but whether the model fits in its memory at all. A model that fits entirely in VRAM runs at the card’s full speed; one that spills over into system RAM slows down sharply, however quick the GPU is on paper.
That is why the Local LLM GPU Comparison weights model fit at 45% of its score, ahead of speed per dollar at 30%, power at 15% and software setup at 10%. It is also why a cheaper 16 GB card can beat a faster 12 GB one, and why a used 24 GB card is often the value pick for anyone willing to accept the heat, power draw and used-card risk. Every speed figure on that page uses the same test — one model, one quantisation level, single-batch — and each one is marked as measured, calibrated from a related test, or estimated.
“Best GPU” depends on the job
The same card can be a smart buy for one build and a poor one for another. A card that is overkill for 1080p gaming may be exactly right for local AI because of its memory, and a flagship that wins every benchmark can still lose on value once price is in the picture. The PC Parts Value Comparator makes the use case the first choice you make, then ranks every card by how well it fits that job and how much fit you get per thousand dollars, with the lowest new and used prices side by side.
Mining: electricity decides
Mining profit is simple arithmetic: daily revenue minus daily electricity cost. Revenue depends on the card’s hashrate and what the coin currently pays; cost is the card’s power draw across 24 hours at your rate per kilowatt-hour. On typical residential rates the gap is thin, and two people with the same card can land on opposite sides of break-even purely because of their power bill. The GPU Mining ROI Calculator takes your own rate and pool fee and shows the result, and the break-even time, before you buy anything.
Where the numbers come from
| Tool | Data it uses | What leaves your browser |
|---|---|---|
| Local LLM GPU Comparison | Published tokens-per-second benchmarks for one fixed test, plus GPU prices from the StashGrid parts feed | Nothing you enter. The page downloads the price file from stashgrid.site. |
| PC Parts Value Comparator | GPU prices from retail listings, refreshed automatically, and a composite benchmark score | Nothing you enter. The page downloads the price file from stashgrid.site. |
| GPU Mining ROI Calculator | Per-coin revenue per unit of hashrate, refreshed automatically | Nothing you enter. The page downloads the coin data file from stashgrid.site. |
Prices are indicative and move daily. Listings that are implausible for a card are rejected rather than shown, and any price that did not come from the live feed is marked on the page. Check the retailer before you buy. Like every StashGrid page, these pages also load Google ads and analytics; the full picture is in the privacy policy.
Questions
How much VRAM do I need to run an LLM locally?
It depends on the model size and how heavily it is quantised. As a rough guide, 8 GB runs small 7B-8B models at 4-bit, 12-16 GB gives room for larger contexts and mid-size models, and 24 GB or more is where 30B-class models become practical. If a model does not fit in VRAM it spills into system memory and slows down sharply, which is why the Local LLM GPU Comparison weights model fit above raw speed.
Why does the same GPU score differently for gaming and for local AI?
Because the bottleneck is different. Gaming at 1080p rarely needs more than 8-12 GB of memory, so raw speed per dollar decides. Local AI is limited first by whether the model fits in VRAM at all, so a slower card with more memory can be the better buy. The PC Parts Value Comparator lets you switch the use case and watch the ranking change.
Are the GPU prices live?
Prices come from a feed of retail listings that StashGrid refreshes automatically, matched on model name and VRAM so a 16 GB card is never priced off its 8 GB sibling. They are indicative only. Implausible listings are rejected, and any price that did not come from the feed is marked on the page. Always check the retailer before buying.
Is GPU mining still profitable?
For most people on residential electricity, only marginally at best. Daily profit is daily revenue minus daily electricity cost, and electricity is usually the deciding number. The GPU Mining ROI Calculator shows daily, monthly and yearly profit and the break-even time for your own electricity rate.
Do these tools send anything I enter to StashGrid?
No. All three calculate in your browser. Each page downloads a price or revenue data file from stashgrid.site when it loads, but nothing you type or select is sent back. Like every StashGrid page, they also load Google ads and analytics, as described in the privacy policy.