当前已收录产品 ¶
本页由构建脚本自动扫描 spi/products/ 目录生成。生成日期：2026-03-25。
当前仓库共收录 1 个产品条目。
产品总表 ¶
| 产品 | 厂商 | 提交方式 | 状态 | 证据等级 | HBD 规模 | 互联 | 源文件 |
|---|---|---|---|---|---|---|---|
| NVIDIA NVL72 | NVIDIA | 仓库内置示例 | 已发布 | 证据支持 | 72 | NVLink 5.0 + NVSwitch 4.0 | spi/products/nvidia/nvl72.yaml |
产品详情 ¶
NVIDIA NVL72¶
| 字段 | 内容 |
|---|---|
| 厂商 | NVIDIA |
| 状态 | 已发布 |
| 提交方式 | 仓库内置示例 |
| 证据等级 | 证据支持 |
| 产品定位 | NVIDIA NVL72 is a rack-scale high-bandwidth domain built around the Blackwell generation. It uses 36 GB200 superchips to form a 72-GPU NVLink domain and represents the current industrial benchmark for vertically integrated SuperPod engineering. |
| 加速器 | 72 |
| 加速器型号 | GB200 / B200 |
| 总 HBD 显存 | 13824 GB |
| 互联技术 | NVLink 5.0 + NVSwitch 4.0 |
| 单跳时延 | 100 ns |
| 域内总带宽 | 130 TB/s |
| 功耗 | 97.2 kW |
| 源文件 | spi/products/nvidia/nvl72.yaml |
委员会评分摘要 ¶
| 维度 | 分数 | 说明 |
|---|---|---|
| 单跳 / 访存延迟 | 5 | 原生内存语义与时延能力 |
| 规模上限 | 4 | 单 HBD 域内规模能力 |
| 拓扑弹性 | 1 | 拓扑重构与调度灵活性 |
| 生态成熟度 | 5 | 软硬件生态与交付成熟度 |
| 功耗与 TCO | 3 | 性能密度与工程成本平衡 |
| 软件复杂度 | 5 | 对用户的软件使用门槛 |
适用负载 ¶
- dense training
- TP-heavy workloads
- large-model inference
雷达图摘要 ¶
NVL72 sits near the Pareto frontier on latency, memory semantics and ecosystem maturity. Its main trade-offs are fixed topology and very high power/cooling requirements, rather than software readiness.
委员会评语 ¶
Example committee assessment based on public materials only. NVL72 is a canonical benchmark product for SPI because it combines advanced packaging, native rack-scale interconnect and mature software into one deliverable system. It should be interpreted as a high-maturity reference point rather than as a vendor-verified SPI submission.