显卡 AI 算力速查：没有显存答案前先查这张表
跑本地大模型，第一道门槛不是算力而是显存——模型装不下就动不了。这张速查表汇总了102 款主流与二手显卡的 AI 算力数据：AI Score、各主流模型的实测 tok/s、显存容量、FP16 算力，按品牌筛选，帮你判断"手里的卡能跑什么模型、速度多快"。
能查到什么
- AI Score 排行：一张图看清从核显到顶级卡的 AI 算力梯队。
- 模型 tok/s 实测：Qwen 系、Llama 系等主流模型的实测输出速度，直接看该卡跑某模型顺不顺。
- 显存与 FP16 算力：判断能装下多大的模型、纯算力够不够。
结合文章看更清楚
选卡逻辑、显存带宽对速度的影响、核显跑模型的可行性，在文章区有多篇实测记录，工具给数字，文章给判断。
GPU AI Compute: check before you buy a card for LLMs
For local LLMs, the first wall is VRAM, not raw compute — if the model doesn't fit, it won't run. This lookup aggregates AI compute data for 102 mainstream and used GPUs: AI Score, measured tok/s on popular models, VRAM size and FP16 compute, filterable by brand, to answer "what can my card run, and how fast".
What you'll find
- AI Score ranking: the whole ladder from iGPUs to flagship cards at a glance.
- Model tok/s: measured output speeds for Qwen, Llama and other popular models.
- VRAM & FP16: judge what fits in memory and whether raw compute is sufficient.
Pair it with articles
For buying logic, bandwidth effects and iGPU feasibility, the articles have hands-on tests — the tool gives numbers, the articles give judgment.