Skip to content

300,000+ products in stockAI infrastructure + enterprise IT hardware. Genuine product, full manufacturer warranty.

GPU Servers hardware: Dell PowerEdge XE9640 Buyer's Guide: Fit, Configuration and Price
GPU Servers
Back to Resources
GPU Servers 6 min read 25 September 2026

Dell PowerEdge XE9640 Buyer's Guide: Fit, Configuration and Price

The XE9640 puts four liquid-cooled H100 SXM GPUs in 2U. It suits racks with direct liquid cooling, and configuration, not list price, sets what you pay.

The Dell PowerEdge XE9640 is the liquid-cooled member of Dell's four-GPU H100 family. It puts four NVIDIA H100 SXM GPUs and two Intel Xeon processors into 2U, and it only makes sense in a rack that can deliver coolant to it. If your facility has direct liquid cooling, it is one of the densest ways to run H100 SXM in a Dell estate. If it does not, the same GPU complex is available air-cooled in the XE8640. This guide covers what the XE9640 is, how it compares with the XE8640, XE9680 and R760xa, and how to get a fair price on one.

What is the Dell PowerEdge XE9640?

  • GPUs: 4x NVIDIA H100 700W SXM, fully connected by NVLink.
  • CPUs: 2x Intel Xeon Scalable, up to 64 cores each.
  • Memory: up to 32 DDR5 DIMMs, up to 2TB at 5600 MT/s.
  • Cooling: direct liquid cooling (DLC) on GPUs and CPUs.
  • Networking: ConnectX-7 NDR InfiniBand ready.
  • Form factor and power: 2U rack, 2800W Titanium hot-swap redundant power supplies.
  • Support: 3-year Dell ProSupport.

The GPUs sit on an SXM baseboard rather than in PCIe slots, so the four H100s exchange data over NVLink directly instead of through the PCIe bus. That is the main technical difference from a PCIe GPU server, and it matters for training and for serving models split across GPUs. Our H100 SXM5 vs PCIe guide covers the trade-off in detail.

Read the configuration in two parts. The GPU complex and the cooling architecture are fixed when you order: you do not convert an XE9640 to air cooling later, and the SXM GPUs are not a field swap to a different generation. CPUs, memory population, NVMe and network adapters are ordinary server choices you can revisit. Spend most of the review time on the first part.

Who is the XE9640 for?

Buyers with direct liquid cooling already in place, or committed to it. Cold plates on the GPUs and CPUs carry most of the heat into a liquid loop, which is what lets four 700W GPUs live in 2U. The rest of the server (memory, drives, network adapters) still relies on chassis airflow, so the room needs air handling as well, just less of it.

It fits organizations standardized on Dell for iDRAC management and a single support contract, and teams that want GPU density per rack without moving to eight-GPU nodes. It does not fit an air-only room. Retrofitting liquid cooling for a handful of nodes means coolant distribution, rack manifolds and facility water work that can outweigh the density gain. Our comparison of DLC and rear-door heat exchangers helps with that decision.

XE9640 vs XE8640 vs XE9680 vs R760xa: which Dell GPU server fits?

  • XE9640: 2U, 4x H100 700W SXM, fully connected NVLink, direct liquid cooling, up to 2TB DDR5. For liquid-cooled racks that want four-GPU density.
  • XE8640: 4U, 4x HGX H100 80GB 700W SXM5, fully connected NVLink, air-cooled (liquid-assisted air cooling with hot-swap fans), up to 2TB DDR5, iDRAC9 Enterprise. The same GPU complex for air-cooled rooms, at twice the rack units.
  • XE9680 8x H100: 6U, 8x H100 80GB SXM5 (640GB HBM3 total), NVLink with NVSwitch at 900GB/s GPU-to-GPU, up to 4TB DDR5 across 32 DIMMs. Also configured with 8x H200 141GB or 8x AMD Instinct MI300X 192GB.
  • R760xa 4x H100 NVL: 2U, 4x H100 NVL 94GB PCIe with NVLink bridges on GPU pairs (up to 600GB/s per pair), up to 4TB DDR5. PCIe only, no SXM. Also available with 4x L40S 48GB for enterprise inference.

XE9640 against XE8640 is a cooling decision, not a performance one: both carry four H100 SXM GPUs on a fully connected NVLink, dual Xeon processors and up to 2TB of memory, so pick by what your racks can cool. XE9640 against XE9680 is a scale decision: if a model needs eight GPUs in one NVLink and NVSwitch domain, read our XE9680 configurations guide. XE9640 against R760xa is an interconnect decision: inference that fits on one or two GPUs runs well on PCIe cards in an air-cooled 2U chassis, with no liquid loop to plan. One more angle for multi-node training: for the same GPU count, four-GPU nodes mean twice as many servers as eight-GPU nodes, and more of the GPU-to-GPU traffic leaves NVLink and crosses the network fabric, so compare the fabric design as well as the server cost.

Can you fill a rack with 2U XE9640 nodes?

Rarely. Liquid cooling moves heat, not power. Four 700W GPUs account for 2.8kW per node before CPUs, memory and fans, so rack power usually runs out long before rack units do. Plan node count from the rack's usable power budget and the cooling loop's capacity; our H100 power and cooling guide walks through the arithmetic. Then confirm with facilities:

  • Coolant distribution unit capacity and the facility water temperature range the platform supports.
  • Rack manifolds and quick-disconnect fittings compatible with the server.
  • Leak detection and the service procedure for disconnecting a node.
  • Per-rack power budget with redundancy, confirmed in writing.
  • Airflow for the components that remain air-cooled.

Where can I buy the Dell PowerEdge XE9640 at the best price?

From an authorized channel, as a configured quote. Two XE9640 quotes can differ widely because they describe different servers. What sets the cost:

  • The GPU complex. Four H100 SXM GPUs are the largest line item and are fixed at order time.
  • CPUs and memory. Core count and DIMM population (up to 2TB) move the total significantly.
  • Networking. One NDR InfiniBand adapter per GPU for multi-node training adds adapters, optics and switch ports; a single-node inference deployment needs far less.
  • Storage. NVMe capacity for datasets, checkpoints and scratch space.
  • Support. ProSupport tier and term. Buying through authorized distribution is what keeps that entitlement valid.
  • Cooling integration. Rack manifolds, coolant distribution and installation are often quoted separately from the server.

Compare quotes line by line against the same configuration. A lower headline figure often means fewer adapters, less memory, a shorter support term or hardware from outside the authorized channel. Ask every supplier to itemize the GPU complex, CPUs, DIMM count and size, drives, network adapters, support term and any cooling hardware, and to confirm the system is new with a valid Dell entitlement. Anything that cannot be itemized is not a like-for-like quote.

With the XE9640 you are choosing a cooling architecture as much as a server. If the rack cannot take liquid, the XE8640 is the same four GPUs in an air-cooled 4U.

Nexus Compute supplies the XE9640, XE8640, XE9680 and R760xa new through authorized distribution, with the full manufacturer warranty. Send the workload, node count, rack power and cooling details and your network fabric, and we return a configured quote within 48 business hours, with purchase orders and net terms available for qualified organizations. Start with a request for quote.

Frequently asked questions

Where can I buy the Dell PowerEdge XE9640 at the best price?

Buy it as a configured system through an authorized channel, so the price reflects the exact GPUs, CPUs, memory, networking, support term and cooling integration you need, and the Dell warranty and ProSupport entitlement are valid. Nexus Compute supplies the XE9640 new and returns configured quotes within 48 business hours. Compare competing quotes line by line against the same configuration.

What GPUs are in the Dell PowerEdge XE9640?

The XE9640 configuration we supply carries four NVIDIA H100 700W SXM GPUs, fully connected by NVLink and cooled by direct liquid cooling. For other accelerator options, ask for a configured quote.

What is the difference between the Dell XE9640 and XE8640?

Both carry four H100 SXM GPUs on a fully connected NVLink, with dual Intel Xeon CPUs and up to 2TB of DDR5. The XE9640 is 2U and direct liquid cooled; the XE8640 is 4U and air-cooled. Choose by what your data center can cool.

Can the XE9640 run in an air-cooled data center?

No. Its GPUs and CPUs are cooled by direct liquid cooling, so it needs a liquid loop with coolant distribution and rack manifolds. For air-cooled rooms, the XE8640 or, for PCIe GPUs, the R760xa are the Dell alternatives.

Should I buy the XE9640 or the XE9680 for LLM training?

If the model and its training state fit across four H100s, or you scale out across nodes over InfiniBand, the XE9640 works and suits liquid-cooled racks. If you need eight GPUs in one NVLink and NVSwitch domain, the 6U XE9680 is the better fit.

How much memory does the Dell PowerEdge XE9640 support?

Up to 32 DDR5 DIMMs and up to 2TB at 5600 MT/s in the configuration we supply. Populate channels evenly across both CPUs and size memory to your data pipeline rather than to the maximum.

Systems covered in this article

Planning a hardware investment?

Tell us what you're trying to build. A procurement specialist will help you specify and quote the right configuration within 48 business hours, no obligation.

Dell PowerEdge XE9640XE9640 priceXE9640 vs XE8640XE9640 vs XE96804x H100 SXM liquid cooled serverDell direct liquid cooling GPU serverDell PowerEdge R760xa H100 NVL