LenovoThinkSystem SR680a V4

The NVIDIA HGX B300 board with eight 288 GB Blackwell Ultra GPUs, two Intel Xeon 6 processors and eight 800 Gb/s ports, in an air-cooled 8U. It's Lenovo's server for training and serving large models without bringing water into the data centre.

Ask for price and lead time See what it supports

Rack 8U

What it comes with

The HGX board is fixed; what you configure is the processors, memory, drives and network. These are its limits according to Lenovo's guide. Lenovo Press product guide

Processor2 Intel Xeon 66700P series · up to 86 cores and 350 W

GPUNVIDIA HGX B3008 × 288 GB HBM3e GPUs · NVLink at 1.8 TB/s

MemoryUp to 4 TB32 DDR5 slots · 6400 MHz with one DIMM per channel

Storage8 × 2.5" NVMePCIe 5.0 · boot from 2 M.2 in RAID

Network8 × 800 Gb/s OSFP portsOnboard ConnectX-8, with GPU Direct

Form factorRack 8UAll connections at the front

Power: Eight Titanium power supplies: 3,200 W in N+1 or 3,800 W in N+N
Cooling: Air: 6 front fans and 15 rear fans for the GPUs
Management: XClarity Controller 3, with the Premier licence included
Operating systems: RHEL 9.6 and Ubuntu 24.04 LTS
Warranty: 3 years, with an on-site engineer next business day

Training and serving very large models

The eight GPUs sit on the same board and talk to each other over NVLink at 1.8 TB/s, so a model split across all of them loses far less time in communication than it would over PCIe. That's what you need to train or fine-tune large models, or to serve a model with hundreds of billions of parameters to a lot of people at once.

The eight 800 Gb/s links with GPU Direct are for joining several SR680a V4 servers into a cluster. A single machine doesn't need them, but if the project is going to grow, it's worth having them from day one.

What we do

We deliver it with RHEL 9.6 or Ubuntu 24.04, the operating systems Lenovo certifies for this model, with the NVIDIA drivers and NCCL. For multi-node training we set up Slurm; for serving models, Kubernetes with the GPU Operator and vLLM. We also leave the 800 Gb/s network and the storage that feeds the GPUs ready, because if storage falls short the eight B300s spend their time waiting for data.

The GPU board

GPU	Memory	How many
NVIDIA B300 (HGX B300)	288 GB HBM3e	8, at 1,100 W each

The HGX B300 board always carries all eight GPUs.

What your data centre needs

This is the one that needs the most planning. It takes 8U, has eight 3,200 W power supplies, or 3,800 W on high-voltage input, and each GPU can draw up to 1,100 W. It's air-cooled with fifteen rear fans just for the GPUs, so the rack needs a lot of power and a cold aisle that can handle that heat. Before anything is agreed we go through your power per rack, PDUs, sockets and cooling with you.

Compared with the rest of the family

Model	What it has	When to choose it
ThinkSystem SR650a V4	Up to 4 RTX PRO 6000 · Intel Xeon 6	The entry point: one to four GPUs in 2U, with Intel.
ThinkSystem SR675 V3	Up to 8 PCIe GPUs · AMD EPYC	Up to eight GPUs with AMD, for several models or teams at once.
ThinkSystem SR680a V3 with B200	NVIDIA HGX B200 · 8 GPUs	The previous HGX, also air-cooled, when lead time or budget comes first.
ThinkSystem SR780a V3	NVIDIA HGX B200 · liquid-cooled	The same eight B200s in 5U, if your data centre has a water loop.
ThinkSystem SR665 V3	4 RTX PRO 4500 · AMD EPYC	Lenovo's entry point for inference, with the Hybrid AI 241 configuration.

See all Lenovo AI and GPU servers

What the delivery includes

From the factory

Server built to your configuration by Lenovo
1- or 3-year warranty depending on the model, with an on-site engineer next business day

From us

GPU, memory and network sizing based on your model
Rack mounting, XClarity and up-to-date firmware
Operating system, NVIDIA drivers and the AI stack
Testing with your workload before it goes into production

If you need it

24×7 on-call support with SLA
Lenovo warranty upgrades: 4- or 2-hour response, or one or two extra years

How we build it

What runs on top of these servers, explained in detail:

RAG over your own documents with Docling AI inference with Ceph, OpenStack and Kubernetes Rocky Linux for AI, with NVIDIA drivers and CUDA ready All our AI services

Support and training for your team

Independent support from the people who installed it Local LLM and GPU infrastructure course Deep learning course with PyTorch and TensorFlow Red Hat OpenShift training

Questions about the ThinkSystem SR680a V4

How does it differ from the SR680a V3 with B200?

Both are air-cooled 8U servers with eight GPUs on an HGX board. The SR680a V4 carries the B300, with 288 GB per GPU instead of 180 GB, Intel Xeon 6 processors instead of 5th Gen Xeon and onboard 800 Gb/s networking. The V3 with B200 is the option when lead time or budget comes first.

Do I need liquid cooling?

No, it's air-cooled, but it concentrates a lot of heat: eight GPUs of up to 1,100 W. If your data centre has a water loop, the SR780a V3 fits eight B200s in 5U with liquid cooling.

Which operating systems does it support?

Lenovo certifies it with Red Hat Enterprise Linux 9.6 and Ubuntu 24.04 LTS. Those are the ones we install, with the NVIDIA drivers and the AI stack you need.

Does your project call for an HGX B300?

Tell us which model you want to train or serve and what data centre you have, and we'll come back with the configuration, what the room needs and a confirmed lead time.

Ask for price and lead time WhatsApp +34 91 198 02 43