GPU comparator for local LLMs

The best GPU for local LLMs is the one whose VRAM holds your model: on prices checked between 29 September 2026 and 30 October 2026, a used RTX 3090 gives 24 GB at 936 GB/s for about $1,350, a new RTX 5090 gives 32 GB at 1,792 GB/s for about $7,681, and a new RTX PRO 6000 Blackwell gives 96 GB at 1,792 GB/s for about $17,063.

Ahmad Osman puts it this way: local AI is capacity × bandwidth × software stack. Capacity is what fits, bandwidth is how fast it can decode, and the stack decides how much of the spec sheet you get to use.

Two questions decide a local AI box: does the model fit, and how much speed do you get for the money. Memory runs up the side and bandwidth per dollar runs across, with bubble size showing raw bandwidth, so up and to the right is best. Cheap gaming cards sit bottom right: good value, too little memory. Big-memory Macs and datacenter cards sit higher but further left: more room, less speed per dollar. Lucebox (yellow) is plotted by its card: 32 GB at 640 GB/s. It also has up to 192 GB of unified memory at 273 GB/s. That is a second, slower memory, so the chart does not add it to the card. Turn Datacenter off to see only what you can buy, and click Lucebox for the build.

up = more memory (what fits)  ·  right = more bandwidth per dollar (speed for the money)  ·  bubble = raw bandwidth  ·  both axes log  ·  tap Lucebox for the build

Sort any column, or tick rows to compare a few. Value is memory × bandwidth ÷ price, so lots of fast memory for the money scores high. Price is the new price while a part is sold and the used price once it is discontinued. For Apple and mini-PCs, memory is the full unified pool. For Lucebox it is the card; the unified pool is a second memory, listed under the name.