按显存需求倒推硬件Work back from memory requirement to hardware

顺序按「先算需要多少显存、再看能不能装下、最后才看算力和功耗」排。每一步只给判断依据，不给推荐结论；算显存用 models 站的算式，本站只回答设备这一端。The order is: work out how much memory is needed, then whether it fits, and only then look at compute and power. Each step gives the criterion, not a recommendation; use the formulas on the models site for the memory sum, since this site only covers the device side.

1. 先确认权重、KV 缓存与框架开销加起来的显存需求，这一步不涉及任何设备。First fix the memory requirement — weights, KV cache and framework overhead together. No device enters the picture yet. 条目entry
2. 需求在 24GB 以内且只要单机：看消费级单卡的显存与整机电源要求。If the requirement is within 24 GB and a single machine is enough, read the consumer card's memory and system power requirement. 条目entry
3. 需求在 48GB 档、且要长时间满载或并排多卡：看 ECC 显存与 300W 档的工作站卡。For the 48 GB tier under sustained load or side-by-side installation, read the workstation card with ECC memory and a 300 W envelope. 条目entry
4. 需求在 80GB 档：确认是 SXM 还是 PCIe，两者的互联能力差一个量级。At the 80 GB tier, establish whether it is SXM or PCIe — their interconnect capability differs by an order of magnitude. 条目entry
5. 需求超过 80GB：比显存容量与带宽，同时看功耗上限是否超出机房供电。Above 80 GB, compare capacity and bandwidth, and check whether the power ceiling exceeds what the facility can supply. 条目entry
6. 单卡装不下又要 8 卡全互联：本地整机看 NVSwitch 拓扑，租用看单实例的卡数与网络。If one card is not enough and eight-way fabric is required, check the NVSwitch topology on-premises or the card count and network per instance in the cloud. 条目entry
7. 要在机器人本体上跑：先看统一内存够不够，再看功耗是否落在电池与散热的范围内。For running on a robot body, first check whether unified memory is sufficient, then whether power fits the battery and thermal envelope. 条目entry
8. 只做视觉推理、不需要大语言模型：低功耗加速模块的算力与功耗档位更贴合。For vision inference only, without a language model, a low-power accelerator module matches the compute and power tier better. 条目entry
9. 最后核供电与场地：整卡功耗、整机电源、机箱风道与工作温度，任何一项不满足都要回头改方案。Finally verify power and site conditions: board power, system supply, chassis airflow and operating temperature. If any one fails, the design goes back a step. 条目entry