Server hardware & datacenter builds
A datacenter build starts with the server. We specify, rack and burn in servers from Supermicro, Dell PowerEdge, HPE ProLiant and Lenovo ThinkSystem, choosing CPUs, memory, storage and GPUs against the workload instead of the catalogue. AMD EPYC 9005 processors, for example, offer up to 192 cores per socket and 12 DDR5 memory channels per socket, which makes memory population a design decision: every channel should carry a DIMM.
What we do
- Server specification by workload: cores, memory channels, NVMe and PCIe layout
- Memory population plans that use every DDR5 channel
- NVMe U.2, U.3 and EDSFF E1.S/E3.S storage layouts
- GPU server builds on HGX-class baseboards with power and cooling budgets
- BMC, IPMI and Redfish management network design
- Firmware and BIOS baselines with change records
- Burn-in with MemTest86, stress-ng and fio before production
- Rack, label and document every server on delivery
Storage has moved from 2.5‑inch U.2 to the EDSFF family. E3.S is designed to replace U.2 and supports x4 to x16 PCIe lanes and up to 70 W; E1.S is the compact form for 1U compute. For AI, the NVIDIA HGX platform puts eight GPUs on one baseboard with NVLink; the H200 carries 141 GB of HBM3e at 4.8 TB/s, and HGX B200 totals 1.4 TB of GPU memory.
Management and proof matter as much as parts. We use Redfish (the DMTF RESTful management standard) and the vendor BMCs, hold firmware to a written baseline, and run MemTest86, stress-ng and fio burn-in before a machine carries production. Experience on these platforms, not a partnership with any vendor. Pairs with datacenter fabric, power and colocation operations.
Leading platforms we work with
- Supermicro
- AMD EPYC
- NVIDIA HGX
- Redfish (DMTF)
- Open Compute Project
- MemTest86
- stress-ng
- fio
- Dell PowerEdge
- Lenovo ThinkSystem
- Intel Xeon 6
- HPE ProLiant
Platforms and practices
- DDR5 RDIMM population
- NVMe U.2, U.3 and E3.S
- EDSFF E1.S
- BMC and IPMI
- Firmware baselines
- OCP-style hardware
References
- Supermicro: server systems families
- Wikipedia: AMD EPYC (9005 Turin: cores, DDR5 channels, SP5)
- NVIDIA H200 Tensor Core GPU (vendor page, preliminary specifications)
- NVIDIA HGX platform (vendor page)
- SNIA: SSD form factors (E1.S, E3.S, U.2)
- Wikipedia: Redfish (specification)
- Open Compute Project: about
- MemTest86
- fio: Flexible I/O Tester
- stress-ng
- Dell: PowerEdge servers (R, XE, XR, T and M series)
- Lenovo Press: ThinkSystem servers (V2, V3, V4; ThinkEdge, ThinkAgile)
- Intel: Xeon processors (Xeon 6 with P-cores and E-cores)
- DMTF: Redfish developer hub (specification, schemas, mockups)
Sources last checked .
Product names are trademarks of their owners and are listed to describe our experience; they do not imply endorsement or partnership.
How we use Server hardware & datacenter builds for clients
Managed IT Services
We run your IT day to day so it stays secure: help desk, devices, Microsoft 365, Exchange, Active Directory, backups, networks and monitoring.
Learn moreCloud Management
Cloud cost review and management, without compromising on quality or security.
Learn moreSuper Intelligence
American‑made Super Intelligence: AI your team can learn, trust and deploy.
Learn moreFrequently asked questions
Direct answers to the questions we hear most. Something else on your mind?
Send us a messageWhich server platforms do you build on?
Supermicro, Dell PowerEdge, HPE ProLiant and Lenovo ThinkSystem, chosen against your workload, not the catalogue. This is experience on these platforms, not a partnership with any vendor.
Why does memory population matter?
Modern processors have many memory channels: AMD EPYC 9005 offers 12 DDR5 channels per socket, and every channel should carry a DIMM. Unpopulated channels cost bandwidth, so population is a design decision.
What storage form factors do you use?
NVMe in U.2, U.3 and the EDSFF family. E3.S is designed to replace U.2 and supports x4 to x16 PCIe lanes; E1.S is the compact form for 1U compute.
Can you build GPU servers for AI?
Yes. Eight-GPU HGX-class systems put eight GPUs on one baseboard with NVLink. We size power, cooling and networking around them, and the memory a model needs decides the GPU choice.
How do you manage firmware and out-of-band access?
We use Redfish and the vendor BMCs and hold firmware to a written baseline, so every machine in a fleet is configured the same way.
What tests run before a server goes into production?
MemTest86, stress-ng and fio burn-in, plus network and thermal checks, accepted against a written checklist.