Frequently Asked Questions
On reading the specs, picking hardware, and how this site works.
Why doesn’t MicrochipMart show prices?
Hardware prices change daily and differ by seller, condition and stock level. Amazon’s Associates terms also restrict displaying their prices to figures sourced live from their Product Advertising API. Rather than publish a number that could be wrong within hours, every listing links straight to Amazon so you see the real, current price there.
How much VRAM do I actually need?
For gaming at 1080p–1440p, 8–12GB is comfortable today. For running large language models locally, VRAM is usually the hard limit — a 7B-parameter model in 4-bit quantization needs roughly 5–6GB just to load, before context; a 13B model needs closer to 10–12GB. If the model doesn’t fit in VRAM, it either runs far slower via system RAM offload or doesn’t run at all. When in doubt for AI work, more VRAM beats a newer GPU generation with less.
What does price per GB of VRAM actually tell me?
It’s a rough efficiency metric: how much you’re paying for raw memory capacity, stripped of everything else the card does. It’s most useful for comparing cards you’d otherwise consider interchangeable — two 16GB cards from different brands, say. It tells you nothing about compute performance, so don’t use it to compare, for instance, a 24GB card against an 8GB card and conclude the 8GB card is the better deal.
CPU cores or GPU — which matters more for AI workloads?
The GPU, almost always. Training and inference for neural networks are dominated by GPU compute and VRAM. The CPU matters for data loading, preprocessing and orchestrating the pipeline — real, but secondary. Don’t stretch a GPU budget thin to afford a bigger CPU.
What’s the difference between a GPU and a dedicated AI accelerator (like Coral or Hailo)?
A GPU is general-purpose: it trains models, renders graphics, and runs inference. A dedicated accelerator like Google Coral or the Hailo-8 is built for one job — fast, low-power inference on a model that’s already trained — typically at the edge, in a Raspberry Pi or embedded device. TOPS (trillions of operations per second) is the relevant spec there, not VRAM.
Are the specifications on this site guaranteed accurate?
They’re drawn from manufacturer documentation and kept up to date as best we can, but board partners vary clocks and configurations, and manufacturers revise specs over a product’s life. Always confirm the exact specification on the retailer’s listing before buying — see our Terms of Use.
Does MicrochipMart get paid for these listings?
Yes — through the Amazon Associates programme. If you buy something after following a link from this site, we may earn a small commission at no extra cost to you. It doesn’t change the listing, the price, or the sort order. See our Affiliate Disclosure for the details.
Which Amazon stores does MicrochipMart link to?
Amazon.com (US) and Amazon.ca (Canada) today. Additional Amazon marketplaces require their own, separate Associates approval, so coverage will expand as those accounts come online.
New vs. used — what should I watch for?
Used GPUs, especially ones pulled from mining rigs, can run hot for years without failing — but check the return window and seller rating before buying, and budget for the possibility of an early failure outside any warranty. New carries a manufacturer warranty and a predictable price, at a premium.