% % %

训练时间的算账：为什么 8-GPU 节点赢在墙钟

LLM 训练的成本可以简化成一道题：同样 1,000 GPU 小时的工作量，用 4 台双卡服务器跑和用 1 台 8 卡服务器跑，墙钟时间差多少？答案不止 4 倍——因为 GPU 间通信是训练的关键路径。PowerEdge XE9680 的 8 颗 GPU 通过 NVLink（NVIDIA）或 Infinity Fabric（AMD）全互联，模型并行（TP/PP）的通信延迟比跨节点走网络低一个数量级。8 卡节点内部通信，4 台双卡节点要跨 100GbE/IB 网络跑同样数据——墙钟时间、能耗、网络带宽全部翻倍。

这就是 XE9680 存在的理由：Dell 首款 8 路 GPU 服务器，专为生成式 AI、ML/DL 训练和 HPC 打造，支持大语言模型、推荐引擎、分子动力学和基因组测序。

Core Specifications

Атрибут	Спецификация
Фактор формы	6U стоечный сервер
Процессор	2 x 5th Gen Intel Xeon Scalable (64 ядра) или 2 x 4th Gen (56 ядра)
Память	32 x DDR5 RDIMM slots, вплоть до 4 ТБ, 5600 МТ/с (5th Gen) / 4800 МТ/с (4th Gen)
Варианты графического процессора	8 x H100 80GB 700W / 8 x H200 141GB 700W / 8 x H20 96GB 500W (all NVLink), 8 x MI300X 192GB 750W (Infinity Fabric), 8 x Gaudi 3 128GB 900W (РОСЕ)
Shared GPU memory	Вплоть до 1.5 TB coherent (MI300X config)
Отсеки для дисков	Вплоть до 8 x 2.5-inch NVMe/SAS/SATA (122.88 ТБ) или 16 x E3.S NVMe
Boot	БОСС-Н1: Аппаратный RAID 1, 2 x M.2 NVMe
PCIe	Вплоть до 10 x Gen5 x16 slots (8 with Gaudi 3)
Сеть	2 x 1GbE embedded + 1 х ОКП 3.0 (х8); 6 x 800GbE OSFP with Gaudi 3
блок питания	3200W Титан (277VAC/260-400VDC) or 2800W Titanium, hot-swap redundant
Управление	iDRAC9, iDRAC Прямой, Redfish RESTful API, OpenManage Enterprise
Размеры / weight	263.2 Икс 482 Икс 1008.77 мм (с безелем); вплоть до 114.05 кг

The GPU Choice: Five Accelerators, One Chassis

The XE9680 is unusual in accepting every major accelerator generation:

Accelerator	Память	Власть	Fabric
NVIDIA H100 SXM5	80 ГБ	700 Вт	НВЛинк
NVIDIA H200 SXM5	141 ГБ	700 Вт	НВЛинк
NVIDIA H20 SXM5	96 ГБ	500 Вт	НВЛинк
AMD MI300X OAM	192 ГБ	750 Вт	Infinity Fabric
Интел Гауди 3 OAM	128 ГБ	900 Вт	Ethernet RoCE (embedded 800GbE)

One chassis, five GPU generations — the platform choice follows the model and the supply line, not a lock-in.

Dell PowerEdge XE9680 6U GPU server — PowerEdge XE9680: 6U 8-GPU training node with NVLink/Infinity Fabric/RoCE accelerator options.

Storage and Expansion for Training Data

Вплоть до 8 x 2.5-inch NVMe/SAS/SATA drives (122.88 ТБ) или 16 x E3.S NVMe drives feed the training pipeline without a separate storage hop; BOSS-N1 handles boot duty in hardware RAID 1. Ten front-facing PCIe Gen5 x16 slots take HBAs, high-speed NICs or NVMe-of-F targets, and the OCP 3.0 слот (х8) adds a 100/400GbE card for multi-node scaling. PERC H965i covers SAS RAID (not supported with Gaudi 3).

Безопасность & Управление

Безопасность. Silicon root of trust, Безопасная загрузка, cryptographically signed firmware, Secured Component Verification, Безопасное стирание, Блокировка системы (iDRAC9 Enterprise/Datacenter), Data at Rest Encryption via SEDs, and TPM 2.0 (FIPS/CC-TCG certified, China variant).

Управление. iDRAC9 with dedicated port, HTML5 virtual console, virtual media and Redfish; OpenManage Enterprise with Power Manager, Service and Update plugins, plus integrations for Ansible, Terraform, ServiceNow and BMC TrueSight — the same fleet plane as every other PowerEdge.

Власть, Охлаждение & Физический

блок питания: 3200W Титан (277 VAC или 260-400 В постоянного тока, US/Canada) or 2800W Titanium (200-240 ВАК / 240 В постоянного тока), hot-swap redundant.
Охлаждение: Air-cooled with up to 6 HPR Gold fans in the mid tray + вплоть до 10 on the rear (12 with Gaudi 3) — designed for 700-900W accelerators.
Физический: 263.2 Икс 482 Икс 1008.77 мм (с безелем), вплоть до 114.05 кг — plan floor load and rail ratings.

ОС & Программное обеспечение

Канонический сервер Ubuntu LTS, Ред Хэт Энтерпрайз Линукс, SUSE Linux Enterprise Server and VMware ESXi; Dell Validated Designs for Generative AI cover the reference workflows end to end.

Часто задаваемые вопросы о покупке

Which GPU? Н200 (141ГБ) for the largest models; H100/H20 for cost tiers; MI300X for AMD estates; Gaudi 3 for Ethernet-only fabrics.
Training or inference? Built for training/HPC; inference workloads may scale better on L40S-class nodes — see our L40S.
Multi-node? 100/400GbE via OCP 3.0; Gaudi 3 builds embed 6 x 800GbE OSFP.
Масса? Вплоть до 114.05 kg fully loaded — rack planning required.
Remote management? iDRAC9 + OpenManage, identical to the rest of the fleet.

Scaling your AI estate? The XE9680 anchors the training tier; xFusion 5885H V7 covers 4-socket compute. As an authorized Dell partner we offer factory-direct pricing, пользовательская конфигурация, 3-year warranty and global shipping — contact us for an AI infrastructure quote.