The hard part of a GPU upgrade isn't picking the fastest card — it's working out whether the money buys anything you'll actually use. This tool nets your current card's resale value off the new one's price, then divides by the VRAM you gain, so you can see what a capacity jump really costs. It also sizes the PSU, projects the extra electricity, and shows the largest model each card holds.
VRAM capacity, not raw speed, is what decides which models you can load at all. Ranked cheapest first:
| GPU | VRAM | MSRP | $/GB | tok/s (8B) |
|---|---|---|---|---|
| Apple M3 Ultra | 512 GB | $3,999 | $8 | 43–90 tok/s |
| Intel Arc B580 | 12 GB | $249 | $21 | 28–57 tok/s |
| AMD Ryzen AI Max+ 395 | 96 GB | $1,999 | $21 | 18–38 tok/s |
| Apple M2 Ultra | 192 GB | $3,999 | $21 | 43–89 tok/s |
| AMD Radeon RX 9060 XT 16GB | 16 GB | $349 | $22 | 24–49 tok/s |
| Intel Arc B570 | 10 GB | $219 | $22 | 24–49 tok/s |
| NVIDIA GeForce RTX 5060 Ti 16GB | 16 GB | $429 | $27 | 40–83 tok/s |
| Apple M4 Max | 128 GB | $3,499 | $27 | 31–64 tok/s |
| Apple M5 Max | 128 GB | $3,499 | $27 | 34–71 tok/s |
| NVIDIA GeForce RTX 3060 (12GB) | 12 GB | $329 | $27 | 26–50 tok/s |
| NVIDIA GeForce RTX 4060 Ti 16GB | 16 GB | $499 | $31 | 31–59 tok/s |
| AMD Radeon RX 7800 XT | 16 GB | $499 | $31 | 39–81 tok/s |
| VRAM | Largest dense model at Q4_K_M | PSU for the fastest card at this tier |
|---|---|---|
| 8 GB | ~11B parameters | 450W (NVIDIA GeForce RTX 5060 Ti 8GB, 180W) |
| 12 GB | ~17B parameters | 550W (NVIDIA GeForce RTX 5070, 250W) |
| 16 GB | ~23B parameters | 750W (NVIDIA GeForce RTX 5080, 360W) |
| 24 GB | ~34B parameters | 850W (NVIDIA GeForce RTX 4090, 450W) |
| 32 GB | ~45B parameters | 1000W (NVIDIA GeForce RTX 5090, 575W) |
| 48 GB | ~68B parameters | 650W (NVIDIA RTX 6000 Ada Generation, 300W) |
| 96 GB | ~135B parameters | 1000W (NVIDIA RTX PRO 6000 Blackwell, 600W) |
| 128 GB | ~180B parameters | 450W (Apple M1 Ultra, 60W) |
Crossing a VRAM tier — 12 GB to 24 GB, or 24 GB to 48 GB — unlocks a whole model class and is usually worth paying for. Adding VRAM within a tier mostly buys context length, and a speed-only upgrade under 50% is hard to notice in day-to-day use. Resale figures shown here are depreciation estimates; check completed listings for a number you can bank on.