Price (trend) is a sparkline of recorded prices, newest on the right — green falling, red rising. Expand a row for the full chart. Current prices are Sept-2026 street figures for new stock where sold new, else typical used; snapshots are appended weekly and on request.
Max model @ Q4 is the largest dense model (billions of params) that loads at ~Q4_K_M with ~2 GB headroom and a modest context: 12 GB → ~14B, 16 GB → ~22B, 24 GB → ~32B, 48 GB → ~70B, 96 GB → ~140B. Longer context eats VRAM fast; MoE and multi-GPU change the math. Pair with the Local AI DB.
AI compute is dense FP16 tensor throughput. Cards without tensor / matrix cores (RDNA2, older Arc, Tesla P40) still run models off raw VRAM but prompt processing is slow and needs ROCm / Vulkan — see the AI note.
Gaming by resolution — Res target is the highest resolution the card drives comfortably at high/ultra without upscaling. The four resolution columns give Cyberpunk 2077 FPS (native, then with DLSS / FSR / Frame Gen) as review composites, ±10–15%. Raster / RT tiers are S (4090 / 5090 class) to D.
Power — TDP, the connector (8-pin vs 12VHPWR / 12V-2x6) and the vendor-recommended PSU. The 8-pin only filter avoids the 16-pin adapter story.