Dell PowerEdge XE9680 vs XE9640
Dell PowerEdge XE9680
Dell PowerEdge XE9640
The choice between the PowerEdge XE9680 and the XE9640 comes down to how many GPUs you want in a single node and whether your data center runs air or liquid cooling. The XE9680 is Dell's flagship 8-way accelerator server, a 6U air-cooled node that puts eight fully NVLink-connected GPUs in one coherent memory domain for the largest training jobs. The XE9640 is a compact 2U, four-GPU node built for direct liquid cooling, designed to pack more accelerated nodes per rack in facilities already plumbed for liquid. Neither is simply better, so the right pick depends on model size, rack density, cooling infrastructure, and budget, and Uniqcli configures and quotes both.
Side by side
| Dell PowerEdge XE9680 | Dell PowerEdge XE9640 | |
|---|---|---|
| Form factor & density | 6U rack node holding a single eight-GPU system; a larger footprint per node and fewer nodes per rack | Compact 2U node with four GPUs; direct liquid cooling lets you stack more accelerated nodes per rack where facility cooling allows |
| GPU configuration & interconnect | Up to eight NVIDIA HGX (H100/H200 class) or AMD Instinct MI300X accelerators, fully connected by NVLink and NVSwitch as one large GPU memory domain | Four NVIDIA SXM GPUs linked by NVLink; a smaller but tightly coupled four-way domain per node |
| Cooling | Air-cooled design that drops into standard air-cooled data halls; liquid-cooled variants of the flagship line exist for the newest accelerator generations | Purpose-built for direct liquid cooling (DLC); requires a facility liquid loop or CDU, and rewards sites already invested in liquid infrastructure |
| Scale-out networking | Abundant PCIe Gen5 expansion for multiple high-speed NICs, enabling GPUDirect RDMA fabrics for very large multi-node training clusters | Fewer slots per node by design; you scale throughput by adding more dense nodes rather than widening a single node |
| Target workloads | Foundation-model and large-language-model training, heavy fine-tuning, and any job that benefits from the biggest single-node pool of GPU memory | Dense inference, mid-size training and fine-tuning, and HPC where liquid density and performance-per-watt matter most |
| Data center requirements | Higher power draw per node but no liquid plumbing needed; the simpler retrofit for existing air-cooled facilities | Needs DLC-ready power and cooling, a higher up-front facility investment that pays back in rack density and thermal efficiency |
| Management & AI Factory | iDRAC and OpenManage with Dell AI Factory validated designs for NVIDIA and AMD accelerators, using air-cooled reference architectures | The same iDRAC and OpenManage tooling with Dell AI Factory liquid-cooled reference designs tuned for high-density deployments |
| Federal & compliance | TAA-compliant configurations available through NASA SEWP V, GSA, and GPC for accredited AI programs | The same TAA-compliant procurement paths; the liquid, high-density option for space- or power-constrained federal data centers |
Shop these now
Live configurations from our catalog with partner pricing. Add to your cart to request a firm quote, or build a full BOM.
Dell PowerEdge XE9680
Dell PowerEdge XE9640
Choose the Dell PowerEdge XE9680 when you need one maximum-GPU node
Pick the XE9680 when the workload is defined by model size and you want the largest coherent GPU domain in a single node. Eight NVLink-connected accelerators with a shared, very large pool of GPU memory make it the workhorse for foundation-model and LLM training and heavy fine-tuning. Its air-cooled design is the simpler fit for data centers that are not yet plumbed for liquid, and its rich PCIe Gen5 networking ties many nodes into a large training fabric. This is the platform when raw single-node capability and standard air cooling lead the decision. Configure a node on /bom and send it to /quote for pricing.
Choose the Dell PowerEdge XE9640 when liquid density and efficiency lead
Pick the XE9640 when your facility already runs direct liquid cooling and you want to pack more accelerated nodes into each rack. The compact 2U, four-GPU design trades single-node GPU count for higher node density and strong performance-per-watt, which suits dense inference, mid-size training, and liquid-cooled HPC. It is the better answer where floor space, power efficiency, and thermal headroom matter more than the biggest possible single node. If you are scaling out by node count rather than widening one node, this is the platform. Spec the cluster on /bom and request pricing at /quote.
Neither server wins outright; they answer two different questions about the same AI build. The XE9680 is the choice when you want the maximum GPUs and GPU memory in one air-cooled node, which makes it ideal for the largest training and fine-tuning jobs in standard data halls. The XE9640 is the choice when you have liquid cooling and want denser, more power-efficient nodes to scale out, which fits inference-heavy and mid-size training estates. Decide based on three things: the size of your models, the density and cooling your facility supports, and how you plan to scale. As an independent integrator and authorized Dell partner, Uniqcli quotes and configures both against your actual workload. Start a build on /bom or reach the team at /contact, and we will model the right node count, networking, and cooling for your AI program.
Talk to a specialistFrequently asked
What is the main difference between the PowerEdge XE9680 and XE9640?
GPU count and cooling. The XE9680 is a 6U, air-cooled node with up to eight NVLink-connected accelerators in a single large GPU domain, built for the biggest training jobs. The XE9640 is a compact 2U, four-GPU node built for direct liquid cooling, designed to scale out dense, efficient nodes in liquid-ready facilities. One maximizes a single node; the other maximizes rack density.
Do I need liquid cooling to run the XE9640?
Yes. The XE9640 is purpose-built for direct liquid cooling, so it expects a facility liquid loop or a coolant distribution unit. If your data center is not yet plumbed for liquid, the air-cooled XE9680 is usually the simpler path, though liquid-cooled variants of the flagship line exist for the newest accelerators. Uniqcli can help assess your facility before you commit.
Which server is better for large-language-model training?
The XE9680 is generally the stronger single-node platform for large-model training because eight fully NVLink-connected GPUs give one very large pool of GPU memory. For clusters, both can scale out over high-speed networking, so the decision also depends on your cooling infrastructure and target rack density. Uniqcli sizes the node count and fabric to your model and budget.
Are both platforms part of the Dell AI Factory?
Yes. Both use iDRAC and OpenManage and are covered by Dell AI Factory validated designs with NVIDIA and AMD accelerators, the XE9680 in air-cooled reference architectures and the XE9640 in liquid-cooled ones. That means tested, repeatable configurations rather than one-off builds.
Can I buy these on a federal contract?
Yes. Both support TAA-compliant configurations and are available through federal procurement paths including NASA SEWP V, GSA, and the Government Purchase Card for smaller orders. Uniqcli can align the build to your accreditation and contract vehicle; start at /quote or /contact.
Build your Dell bill of materials.
Send us the requirement, the project, or an existing quote to beat. We come back with a validated, TAA-compliant Dell configuration and a real price, often below list.
[email protected] · Chicago, IL