GPU 目录与可用容量
Browse hardware the calculator understands before running a workload-specific recommendation. Catalogue entries are separate from live capacity and approved pricing.
NVIDIA · Ada Lovelace
NVIDIA L40S 48GB
适合中型 LLM、图像生成、计算机视觉和成本可控的生产推理。
- 显存
- 48 GB
- 内存
- GDDR6 ECC
- 带宽
- 864 GB/s
NVIDIA · Ada Lovelace
NVIDIA L4 24GB
适合小模型推理、embedding、reranking、开发和成本敏感 AI 服务。
- 显存
- 24 GB
- 内存
- GDDR6
- 带宽
- 300 GB/s
从已验证 GPU 库存中选择
先用目录选择硬件；当模型规模、上下文或预算很关键时，再使用推荐器。
GPU 规模估算方式
计算器会先估算模型权重、KV 缓存、运行时工作区、工作负载开销和生产余量，再筛选库存。
推理与训练
推理和训练分别估算。训练会包含优化器状态、梯度或适配器、激活、运行空间和拓扑限制。
精度与量化
较低精度可以减少权重内存，但量化元数据、运行空间和技术限制仍然需要考虑。
版本化报价逻辑
报价请求会绑定模型快照、计算器版本、所选配置和价格版本。
AI GPU hosting 常见问题
How is VRAM estimated?
The estimator combines model weights, KV cache, runtime workspace, batching, workload overhead and safety headroom.
Is training sized differently from inference?
Yes. Fine-tuning and training include gradients, optimiser state, activations and topology constraints.
Can I use a custom model?
Yes. You can enter parameter count and architecture details; lower-confidence estimates are labelled.
Are recommendations benchmarks?
No. Estimated results and measured benchmark data are labelled separately.
Can unavailable GPUs be purchased?
No. Unavailable hardware can remain in the catalogue, but it is not offered as immediately purchasable inventory.
How does pricing work?
Approved prices are shown where cost and market inputs have been reviewed. Otherwise the offer remains request-quote.