Glossary
Server hardware, storage and import terms, defined
The vocabulary you meet when speccing and importing enterprise servers, written to be read once and understood. Each entry answers the question first, then adds the detail that matters when you are sourcing hardware from China.
Remote management
- iDRAC: Integrated Dell Remote Access Controller
- iDRAC is Dell's dedicated out-of-band management controller built into PowerEdge servers. It runs on its own processor and network port, letting admins power-cycle, reinstall, and monitor a server remotely even when the OS is down or the machine is powered off.
- HPE iLO: Integrated Lights-Out
- iLO is HPE's embedded out-of-band management controller for ProLiant servers. It runs on a separate processor and network port, giving remote power control, virtual console, virtual media, and hardware health monitoring even when the host OS is offline.
- BMC: Baseboard Management Controller
- A BMC is a small dedicated processor on a server motherboard that manages the hardware independently of the main CPU and OS. It monitors sensors, controls power, and provides remote access. iDRAC, iLO, and Huawei iBMC are vendor implementations of a BMC.
- IPMI: Intelligent Platform Management Interface
- IPMI is a standardized hardware management interface that lets administrators monitor and control a server through its BMC, independently of the CPU and OS. It handles sensor readings, power control, and remote console over LAN, but is now considered legacy in favor of Redfish.
- Redfish: DMTF management API
- Redfish is a modern RESTful API standard from the DMTF for managing servers and datacenter hardware. It uses HTTPS and JSON to expose inventory, sensors, power control, and firmware updates, and is the intended replacement for the older IPMI protocol.
Memory
- RDIMM: Registered DIMM
- An RDIMM is a server memory module with a register (buffer) between the memory controller and the DRAM chips for address and command signals. Buffering reduces electrical load on the controller, allowing more DIMMs per channel and greater stability than unbuffered UDIMMs.
- LRDIMM: Load-Reduced DIMM
- An LRDIMM is a server memory module that buffers both data and address/command signals through an on-module memory buffer. This cuts electrical load further than an RDIMM, enabling the highest per-DIMM capacities and full memory population, at slightly higher latency.
- ECC memory: Error-Correcting Code
- ECC memory is RAM that detects and corrects single-bit errors and detects multi-bit errors using extra parity bits. It prevents silent data corruption and crashes from cosmic-ray bit flips, which is why it is standard in servers and required for reliable 24/7 operation.
- Memory rank: single / dual / quad rank
- A memory rank is a 64-bit (72-bit with ECC) block of DRAM chips accessed together by the memory controller. A DIMM can have one, two, or four ranks. Rank count affects capacity, performance through interleaving, and how many DIMMs a channel can support.
- Memory channels & DIMM population
- A memory channel is an independent path between the CPU and its DIMMs. Total bandwidth scales with the number of populated channels, so filling every channel matters more than filling every slot. Modern Xeon and EPYC CPUs have 8 or more channels per socket.
- DDR4 vs DDR5 for servers
- DDR4 server memory runs up to 3200 MT/s and is used by Xeon Scalable 1st–3rd gen and EPYC through Milan. DDR5 starts at 4800 MT/s, reaches 5600 MT/s and higher, and is required by Xeon Sapphire Rapids and later and by EPYC Genoa. The two are not interchangeable.
CPU
- Intel Xeon Scalable generations (1st–5th gen)
- Intel Xeon Scalable spans five generations: 1st gen Skylake, 2nd gen Cascade Lake, 3rd gen Ice Lake, 4th gen Sapphire Rapids, and 5th gen Emerald Rapids. Each step raised core counts and memory speed, with the shift from DDR4 to DDR5 arriving at the 4th generation.
- AMD EPYC generations (Naples to Genoa/Bergamo)
- AMD EPYC spans Naples (1st gen), Rome (2nd gen), Milan (3rd gen), and Genoa/Bergamo (4th gen). Rome and Milan use socket SP3 with 8 DDR4 channels and up to 64 cores; Genoa moved to socket SP5 with 12 DDR5 channels, PCIe 5.0, and up to 96 cores.
- TDP: Thermal Design Power
- TDP is the sustained heat output in watts that a CPU's cooling must dissipate at its base operating point. It guides heatsink, fan, and rack cooling design. A higher TDP generally means more cores or higher clocks, and more heat and power draw to plan for.
Procurement
- CTO: Configure-to-Order
- CTO means a server is built to the buyer's chosen specification of CPU, memory, storage, and cards rather than sold as a fixed prebuilt model. It lets you match hardware exactly to a workload, versus off-the-shelf SKUs with predetermined components.
- Refurbished vs recertified vs new servers
- New servers are factory-fresh with full warranty. Refurbished servers are used units tested and returned to working order, often by a third party. Recertified servers are refurbished to a defined standard, frequently by or on behalf of the OEM, with a documented warranty. Price rises from refurbished to recertified to new.
Storage
- NVMe: Non-Volatile Memory Express
- NVMe is a storage protocol that lets SSDs talk directly to the CPU over PCIe lanes instead of the older SATA or SAS controller path. It cuts latency and pushes millions of IOPS, far beyond what SATA SSDs deliver.
- SAS vs SATA drives
- SAS and SATA are two drive interfaces. SAS-3 runs at 12 Gbps with dual ports, full-duplex signaling, and better error handling for enterprise arrays; SATA III runs at 6 Gbps, single port, half-duplex, and costs less per gigabyte.
- U.2 and U.3 drive form factors
- U.2 and U.3 are 2.5-inch enterprise SSD form factors that carry NVMe over PCIe. U.3 uses the newer SFF-TA-1001 connector so one tri-mode backplane bay accepts NVMe, SAS, or SATA; U.2 bays are wired for NVMe or SATA only.
- RAID levels (0, 1, 5, 6, 10)
- RAID combines drives for speed, redundancy, or both. RAID 0 stripes with no protection, RAID 1 mirrors, RAID 5 uses single parity (survives one drive loss), RAID 6 uses double parity (survives two), and RAID 10 mirrors then stripes for speed plus redundancy.
- Hardware RAID vs software RAID
- Hardware RAID runs on a dedicated controller with its own processor and cache, offloading parity from the CPU and often adding battery-backed write cache. Software RAID (mdadm, ZFS, Storage Spaces) runs on the host CPU with no extra card and no vendor lock to a specific controller.
- DWPD: Drive Writes Per Day (SSD endurance)
- DWPD measures how many times an SSD's full capacity can be overwritten every day for its warranty period, usually five years. A 1.92 TB drive rated at 3 DWPD tolerates about 5.76 TB of writes per day. Higher DWPD means higher write endurance.
Hardware
- Hot-swap (drives, PSU, fans)
- Hot-swap means you can remove and replace a component while the server keeps running, with no shutdown. It applies to drives in a backplane, redundant power supplies, and fan modules, and requires a chassis and backplane designed for live insertion.
- Rack unit (U) and 1U/2U/4U form factors
- A rack unit (U) is 1.75 inches (44.45 mm) of vertical space in a standard 19-inch rack. A 1U server is one unit tall, a 2U is two, a 4U is four. Taller chassis fit more drives, larger coolers, and full-height GPUs at the cost of rack density.
- Redundant power supply (1+1 / N+1)
- A redundant power supply uses two or more hot-swap PSUs so the server keeps running if one fails or loses its feed. 1+1 means one spare backs one active unit; N+1 means one spare backs N active units across a larger system. Both allow live replacement without downtime.
- PCIe lanes and generations (Gen3/4/5)
- PCIe moves data over lanes, and each generation doubles per-lane bandwidth. Gen3 is about 1 GB/s per lane per direction, Gen4 about 2 GB/s, Gen5 about 4 GB/s. A Gen4 x16 slot delivers roughly 32 GB/s per direction; Gen5 x16 doubles that to about 64 GB/s.
GPU
- SXM vs PCIe GPUs (data-center)
- SXM is NVIDIA's socketed GPU module that mounts on a baseboard with NVLink between GPUs, higher power limits, and better cooling. PCIe GPUs are standard cards in expansion slots. For the same generation, SXM parts run at higher TDP and deliver faster GPU-to-GPU bandwidth than PCIe.
- NVLink: NVIDIA GPU interconnect
- NVLink is NVIDIA's high-speed interconnect that links GPUs directly, bypassing the PCIe bus for far greater GPU-to-GPU bandwidth. Combined with NVSwitch, it lets every GPU in a node share memory access at bandwidth PCIe cannot reach, which speeds large-model training.
- HBM: High-Bandwidth Memory
- HBM is stacked DRAM placed next to the GPU die on the same package and connected by a very wide interface, giving several terabytes per second of memory bandwidth. Data-center accelerators use HBM instead of GDDR because model weights and activations demand extreme bandwidth.
Logistics
- Incoterms for hardware (EXW, FOB, DAP, DDP)
- Incoterms define who handles shipping, risk, and customs at each step. EXW: buyer collects from the seller's door and does everything after. FOB: seller clears export and loads at the origin port. DAP: seller delivers to destination, buyer clears import and pays duties. DDP: seller delivers duty-paid.
Speccing a build and not sure which of these you need? Send us the workload and we map it to the right CPU, memory, storage and GPU, then send an EXW Hangzhou quote within 1 hour during business hours (Mon–Fri 09:00–19:00 Beijing time).