Knowledge Base

The standard measurement for storage performance. It represents how many read or write requests a drive can handle in a single second. "Functional" IOPS — what an array delivers after RAID overhead — is almost always lower than a drive's rated burst IOPS.

A method of combining multiple physical drives into a logical volume for performance, fault tolerance, or both. RAID 1 mirrors data across drives; RAID 5 and 6 stripe data with parity to survive one or two drive failures; RAID 10 stripes mirrored pairs for both speed and redundancy.

The number of physical I/O operations required to complete a single logical write, caused by parity or mirroring overhead. RAID 1/10 has a penalty of ×2, RAID 5 is ×4, and RAID 6 is ×6 — which is why write-heavy workloads lose more usable IOPS on higher-parity RAID levels.

RPO (Recovery Point Objective) is the maximum acceptable amount of data loss, measured in time since the last good backup. RTO (Recovery Time Objective) is the maximum acceptable downtime before systems must be back online. An RPO of 4 hours and RTO of 2 hours means you can lose up to 4 hours of data and must be operational within 2 hours of declaring a disaster.

Drive manufacturers advertise capacity in base-10 (1 TB = 1,000,000,000,000 bytes), while operating systems report capacity in base-2 (1 TiB = 1,099,511,627,776 bytes). A "12TB" drive will format to roughly 10.9 TiB in your OS — that's a naming convention difference, not missing capacity.

A storage protocol built for flash media that connects directly over PCIe instead of a legacy SATA/SAS controller. NVMe drives dramatically cut command latency and can deliver anywhere from tens of thousands to over a million IOPS, versus low hundreds for spinning disk.

A notation for describing an IP network's size using a prefix length, such as /24. The prefix indicates how many bits are fixed for the network portion of the address — the remaining bits determine how many hosts the subnet can hold.

A 32-bit value that splits an IP address into network and host portions. It's the dotted-decimal equivalent of a CIDR prefix — for example, /24 and 255.255.255.0 describe the same subnet boundary.

Address ranges reserved for internal networks and non-routable on the public internet: 10.0.0.0/8, 172.16.0.0/12, and 192.168.0.0/16. Any organization can reuse these ranges internally without conflicting with the public internet.

The highest address in a subnet, reserved for sending a message to every host on that network segment. Along with the network address (the lowest address), it's one of the two addresses excluded from a subnet's usable host count.

A vCPU is a virtual CPU core allocated to a VM, backed by a slice of a physical core's scheduling time. Overcommitting assigns more vCPUs than physical threads exist — a 4:1 ratio is typical for general enterprise workloads, while latency-sensitive apps should stay closer to 2:1.

The maximum amount of heat, in watts, a CPU or GPU is designed to dissipate under sustained load. TDP is a thermal ceiling set by the vendor, not a measurement of real-world power draw — actual draw typically runs 50–80% of TDP depending on workload.

Dedicated memory on a GPU used to hold model weights, activation data, and the KV cache during inference. Sizing VRAM for an LLM means accounting for both the model's weights (parameters × bytes-per-parameter) and per-user context memory, plus overhead for CUDA kernels.

A measure of how much work a CPU architecture completes per clock cycle. Two chips at the same clock speed can have very different real-world throughput if their IPC differs — which is why raw GHz is a misleading way to compare enterprise processors.

Reducing the numeric precision used to store a model's parameters to shrink memory footprint. FP16 (2 bytes/param) preserves full quality; INT8 (1 byte/param) trims VRAM in half with minimal quality loss; INT4 (0.5 bytes/param) is aggressive quantization used when memory is the primary constraint.

The ratio of total facility power to power delivered to IT equipment (Total Facility W ÷ IT Load W). A PUE of 1.0 is a theoretical perfect score; 1.2 is excellent hyperscale-grade efficiency; 1.5 is a typical enterprise data center; 2.0+ signals an aging or inefficient facility.

kW (kilowatts) is real, usable power; kVA (kilovolt-amps) is apparent power, which includes reactive power lost to inefficiency. The ratio between them is the power factor — modern server PSUs typically run a power factor near 0.95–0.99, so the two figures are close but rarely identical.

For loads running longer than 3 hours (which servers always are), NEC Article 210.20 requires that the sustained load not exceed 80% of a branch circuit's rating. A 20A circuit is therefore only safe to load to 16A continuously.

The standard unit of vertical rack space, defined by EIA-310 as 1.75 inches (44.45mm). A standard full-height rack is 42U; a 2U server occupies 3.5 inches; GPU chassis and storage heads commonly need 4U or more.