Glossary — HPC / GPU acronyms decoded

Working operator's reference for the soup of acronyms in GPU/HPC infrastructure. Optimized for the question 'what does this letter-soup actually mean and where does it sit in the stack'.

▶Try the commands on this page in the command emulator — type help for the full list, or solutions for copy-paste fix recipes.

This page is a flat alphabetical index of the acronyms you trip over when reading runbooks, vendor docs, and incident reports. Each entry: what the letters stand for, the one-line meaning, and where to read more.

A

  • ABI — Application Binary Interface. The contract between a compiled binary and the kernel/libraries it links to. Relevant when CUDA toolkit or NCCL versions are mismatched between container and host.
  • ACS — Access Control Services. PCIe security feature that, if enabled on bridges between GPU and NIC, breaks GPUDirect-RDMA peer-to-peer transfers. See /docs/kernel-tuning/acs.
  • AER — Advanced Error Reporting. PCIe-level error logging in the kernel; reads via dmesg | grep AER. Escalating AER errors usually precede a GPU fell off the bus event.
  • AMX — Advanced Matrix Extensions (Intel). CPU instruction set for tensor ops. Relevant for inference frontends running on Sapphire Rapids+ Xeon.
  • AOC — Active Optical Cable. Optical IB/Ethernet cable with built-in transceivers, used for runs longer than DAC can do (3m+).
  • API server — kube-apiserver. The K8s control plane HTTP service that everything talks to.
  • ARP — Address Resolution Protocol. IP→MAC mapping; rarely matters for IB but matters for IPoIB and RoCE management traffic.
  • AVX-512 — Advanced Vector Extensions 512-bit. CPU SIMD; Slurm/MPI workloads benefit.

B

  • BAR — Base Address Register. PCIe memory window. BAR1 on H100 = 64 GB; on B200 = 256 GB. Sized to be large enough for GDR.
  • BDF — Bus:Device.Function. PCIe address format like 0000:18:00.0. Read via lspci.
  • BER — Bit Error Rate. Physical layer error rate on a network link. IB target ≤ 1E-12; values above 1E-9 = bad cable/transceiver. See /docs/networking/ib-switches-l2.
  • BFB — BlueField Boot. The DPU image bundle. Flashed onto the BlueField management filesystem.
  • BlueField — NVIDIA's DPU (data processing unit) family. Smart NIC + ARM + accelerators.
  • BMC — Baseboard Management Controller. The out-of-band management micro running on the motherboard. Reads chassis sensors, power-cycles.
  • BPF / eBPF — Berkeley Packet Filter (extended). Kernel verifier-checked programs for tracing, networking (Cilium), security (cgroup device filtering).
  • busbw — Bus Bandwidth (NCCL terminology). The derived bandwidth at the underlying interconnect, accounting for the algorithm. The honest cross-topology comparison metric. See /docs/operations/nccl-tests.

C

  • cBPF — classic BPF. The older, classic version of BPF (vs eBPF).
  • CC — Confidential Computing. Memory encryption between GPU and host (H100+). Separate domain from "congestion control" abbrev.
  • CDI — Container Device Interface. The modern way to expose /dev/nvidia* and CUDA libs into containers (replaces legacy --gpus all). See /docs/drivers/container-toolkit.
  • CDNA — Compute DNA. AMD's HPC GPU architecture (MI100/MI200/MI300 series).
  • CDU — Coolant Distribution Unit. Plumbing for direct-to-chip liquid cooling.
  • CRD — Custom Resource Definition. Kubernetes way to define new API objects (e.g., SlurmCluster, Reservation).
  • CRI — Container Runtime Interface. K8s ↔ containerd/cri-o boundary.
  • CRC — Cyclic Redundancy Check. Error-detecting code; NVLink CRC errors = wire corruption.
  • CUDA — Compute Unified Device Architecture. NVIDIA's GPU programming model + toolkit.
  • cuDNN — CUDA Deep Neural Network library. GPU primitives for ML frameworks.

D

  • DAC — Direct Attach Copper. Short copper cable with built-in connectors; cheaper than AOC up to ~3m.
  • DCB — Data Center Bridging. Ethernet extensions (PFC, ETS, DCBX) needed for lossless RoCE.
  • DCGM — Datacenter GPU Manager (NVIDIA). Daemon + CLI (dcgmi) + Prometheus exporter for GPU monitoring/health. See /docs/nvidia/dcgm.
  • DCQCN — Datacenter Quantized Congestion Notification. Mellanox congestion control on RoCE; reacts to ECN marks. See /docs/networking/roce.
  • DGX — NVIDIA's reference AI workstations/servers (DGX-1, DGX-2, DGX A100, DGX H100, DGX H200, DGX B200, DGX GB200).
  • DKMS — Dynamic Kernel Module Support. Auto-rebuilds kernel modules (nvidia.ko, peermem) when kernel updates.
  • DLC — Direct Liquid Cooling. Cold plate on the GPU/CPU; the dominant cooling for H100/B200 dense racks.
  • DMA — Direct Memory Access. Hardware-driven memory transfer without CPU involvement.
  • DOCA — Datacenter-on-a-Chip Architecture. NVIDIA SDK for programming BlueField DPUs.
  • DPDK — Data Plane Development Kit. Userspace packet processing framework; Weka uses it for high-perf network IO.
  • DPU — Data Processing Unit. SmartNIC + ARM cores (BlueField is one).
  • DRA — Dynamic Resource Allocation. K8s alpha API for fine-grained GPU/device claims.

E

  • EC — Erasure Coding. Storage data protection scheme. Weka uses configurable EC schemes (e.g., 4+2, 8+2, 16+4).
  • ECC — Error-Correcting Code. Memory hardware error detection/correction. SBE = single-bit (corrected); DBE = double-bit (uncorrectable).
  • ECN — Explicit Congestion Notification. IP/TCP flag (and IB CNP) for backoff signaling. Critical for RoCE DCQCN.
  • EDR — Enhanced Data Rate. IB speed (100 Gbps lane × 4 = 400 Gbps); succeeded by HDR, NDR.
  • EVPN — Ethernet VPN. L2/L3 fabric overlay; less common in pure HPC fabrics but appears in mixed environments.

F

  • FE / BE — Frontend / Backend. Common in storage architecture (Weka FE clients vs BE storage hosts).
  • FEC — Forward Error Correction. Recovery codes added to physical-layer transmission. RS-FEC standard for NDR.
  • FM / fabric-manager — NVIDIA service that coordinates NVSwitch fabric. Must match nvidia-driver minor version.
  • FP4 / FP6 / FP8 / FP16 / FP32 / FP64 — Floating point with N bits. Newer GPUs add lower-precision formats (FP8 on H100, FP4 on B200) for AI throughput.
  • FSDP — Fully Sharded Data Parallel (PyTorch). Distributed training strategy.

G

  • GDR — GPUDirect-RDMA. Zero-copy data path GPU→NIC→wire. Requires nvidia-peermem + ACS off + IOMMU pt. See /docs/networking/gpudirect.
  • GDS — GPUDirect Storage. Zero-copy from storage to GPU memory.
  • GFD — GPU Feature Discovery. K8s component that labels nodes with GPU details (used by GPU Operator).
  • GID — Global Identifier. IB address (port-level GUID-derived); used in RoCE for IP-equivalent routing.
  • GPU Operator — NVIDIA Helm-deployed K8s operator that manages driver, container-toolkit, device-plugin, DCGM, MIG. See /docs/kubernetes/gpu-operator.
  • GRES — Generic Resources (Slurm). Used to declare GPUs/network/etc per node.
  • GUID — Globally Unique Identifier. 64-bit IB address baked into HCA firmware.

H

  • HBM — High Bandwidth Memory. Stacked DRAM glued to the GPU package. HBM2 → HBM2e → HBM3 → HBM3e (each generation = more capacity + bandwidth).
  • HCA — Host Channel Adapter. The IB equivalent of a NIC. ConnectX cards.
  • HDR — High Data Rate. IB speed (50 Gbps lane × 4 = 200 Gbps). Succeeded NDR.
  • HGX — NVIDIA reference baseboard for OEM AI servers (Supermicro, Foxconn, etc).
  • HOL — Head of Line (blocking). Network congestion pattern where one stuck packet blocks others. PFC pause storms cause HOL.
  • HPC — High Performance Computing. Traditional compute-heavy scientific workloads + modern AI training.
  • HWE — Hardware Enablement (Ubuntu). Newer kernel + Mesa stack on LTS releases.

I

  • IB — InfiniBand. Layer-2-only fabric. See /docs/networking/infiniband.
  • IOMMU — IO Memory Management Unit. Translates device DMA addresses. iommu=pt (pass-through) is the recommended mode for HPC; full virtualization adds latency. See /docs/networking/pcie-topology.
  • IPMI — Intelligent Platform Management Interface. Out-of-band server management protocol; mostly superseded by Redfish.
  • IPoIB — IP over InfiniBand. IP layer on top of IB; not used by NCCL (which goes verbs direct).
  • IRQ — Interrupt Request. Hardware → CPU notification. Network/GPU IRQ pinning matters for NUMA-aware perf.
  • ISA — Instruction Set Architecture. CPU x86_64, ARM64, etc.
  • ISIS / IS-IS — Intermediate System to Intermediate System. Routing protocol; rare in HPC fabrics, common in IP backbones.

J

  • Jumbo frame — Ethernet frame > 1500 bytes (typically 9000). Mandatory for high-bandwidth Ethernet. IB MTU 4096 is separate.

K

  • K8s — Kubernetes. Container orchestrator.
  • KV cache — Key-Value cache (LLM inference). Stores attention key/value tensors per token; sized like the model.
  • kubelet — K8s node agent. Runs on every node, talks to API server, manages pods.

L

  • LDAP — Lightweight Directory Access Protocol. Identity provider (Authentik exposes LDAP outpost).
  • LID — Local Identifier. IB L2 address assigned by SM; analogous to MAC. Max ~49152 LIDs per subnet.
  • LMC — LID Mask Count. Number of LIDs per port (for path diversity).
  • LTSB / LTS — Long Term Support Branch. NVIDIA driver branch with extended support (R535, R570 historically).

M

  • MAC — Media Access Control. Ethernet L2 address.
  • MCM — Multi-Chip Module. Package with multiple silicon dies (GH200, GB200 superchip).
  • MFT — Mellanox Firmware Tools. mst, flint, mlxlink, mlxconfig. See /docs/drivers/ofed.
  • MIG — Multi-Instance GPU. Hardware partition of one GPU into up to 7 isolated instances. See /docs/nvidia/mig.
  • MLAG — Multi-Chassis Link Aggregation. Ethernet HA for switches.
  • MLNX_OFED — Mellanox OpenFabrics Enterprise Distribution. Vendor-packaged OFED stack.
  • MOFED — synonym for MLNX_OFED.
  • MPI — Message Passing Interface. Distributed compute API; OpenMPI / MPICH / MVAPICH implementations.
  • MPS — Multi-Process Service (NVIDIA). Lets multiple processes share a single GPU; alternative to MIG.
  • MR — Memory Region (RDMA). Registered memory addressable via verbs.
  • MTU — Maximum Transmission Unit. Ethernet frame size; IB has its own MTU concept (256/512/1024/2048/4096 bytes).

N

  • NAT — Network Address Translation. Used in some bare-metal providers' shared-IP egress. Often the bandwidth bottleneck on egress-heavy workloads.
  • NCCL — NVIDIA Collective Communications Library. Implements all_reduce, all_gather, etc. across GPUs. See /docs/networking/nccl.
  • NDR — Next Data Rate. IB speed (100 Gbps lane × 4 = 400 Gbps). Current production class.
  • NeMo — NVIDIA framework for training/fine-tuning LLMs/speech/multimodal.
  • NFS — Network File System. Older shared FS protocol; rarely sufficient for HPC GPU workloads.
  • NIC — Network Interface Card. The Ethernet equivalent of HCA.
  • NIM — NVIDIA Inference Microservices. Pre-packaged inference containers (Llama, Mistral).
  • NIXL — NVIDIA Inference Xfer Library. Point-to-point tensor transport for distributed inference (vs NCCL's collective focus). See /docs/nvidia/nixl.
  • NPU — Neural Processing Unit. Generic term for AI-specific accelerators.
  • NUMA — Non-Uniform Memory Access. Multi-socket CPU memory topology; pinning to local NUMA matters for performance. See /docs/kernel-tuning/numa.
  • NVL — NVLink (e.g., NV18 = 18 NVLinks per GPU on H100; NVL72 = 72-GPU rack-level fabric on Blackwell).
  • NVLink — NVIDIA proprietary GPU-to-GPU interconnect (1.8 TB/s per H100).
  • NVMe — Non-Volatile Memory Express. Storage protocol over PCIe.
  • NVSHMEM — NVIDIA SHared MEMory. Cross-GPU shared memory abstraction.
  • NVSwitch — NVIDIA on-baseboard switch creating a flat NVLink mesh across the 8 GPUs of a node (and across racks in NVL72).

O

  • OCP — Open Compute Project. Hardware open spec consortium.
  • OFED — OpenFabrics Enterprise Distribution. Linux RDMA software stack.
  • OOM — Out Of Memory. Linux kernel kills processes; cgroup-scoped OOM kills offending pod.
  • OpenSM — Reference open-source IB Subnet Manager. Often runs embedded on the switch instead.
  • OSFP — Octal Small Form-factor Pluggable. Connector for 400/800G IB/Ethernet.

P

  • PAM — Pluggable Authentication Modules. Linux auth stack; Slurm uses pam_slurm_adopt.
  • PCIe — Peripheral Component Interconnect Express. Generation/lanes table: gen3 ~1 GB/s/lane, gen4 ~2 GB/s/lane, gen5 ~4 GB/s/lane, gen6 ~8 GB/s/lane.
  • PD — Protection Domain (RDMA). Resource grouping for QPs/MRs.
  • peermem — nvidia-peermem kernel module. Lets the IB stack DMA into GPU VRAM. Required for GDR.
  • PFC — Priority Flow Control (IEEE 802.1Qbb). Per-traffic-class pause frames; foundation of lossless RoCE.
  • PKey — Partition Key (IB). Tag for L2 isolation; used for multi-tenant separation.
  • PMU — Performance Monitoring Unit. CPU hardware counters for profiling.
  • PSU — Power Supply Unit.
  • PUE — Power Usage Effectiveness. Total facility power / IT power. AI factories target 1.15–1.30 with liquid cooling.
  • pyxis — SPANK plugin that integrates enroot with Slurm.

Q

  • QoS — Quality of Service. Used both in Slurm (job priority/limits) and networking (DSCP/PFC priority).
  • QP — Queue Pair (RDMA). Send + Receive queue per RDMA connection. Reliable Connected (RC), Unreliable Datagram (UD), etc.
  • QSFP — Quad SFP. Connector form factor for 100G+ networking. QSFP28, QSFP-DD.

R

  • RBAC — Role-Based Access Control. K8s API authz layer.
  • RC / UC / UD / RD — Reliable Connected / Unreliable Connected / Unreliable Datagram / Reliable Datagram. RDMA QP types. NCCL uses RC.
  • RCCL — ROCm Collective Communications Library. AMD's NCCL fork.
  • RDMA — Remote Direct Memory Access. Zero-copy network protocol; the foundation of NCCL multi-node performance.
  • RKE2 — Rancher Kubernetes Engine 2. Fully-conformant K8s distro, common on bare metal.
  • RoCE — RDMA over Converged Ethernet. v1 = L2-only; v2 = UDP-encapsulated, routable. Requires lossless (DCB+PFC).
  • ROCm — Radeon Open Compute. AMD's software stack.
  • RSS — Receive Side Scaling. NIC feature spreading received packets across CPU queues.

S

  • sacctmgr — Slurm DB management CLI. Add/modify accounts, users, QoS.
  • sbatch / srun / scontrol — Slurm CLI commands (submit batch, run interactive, control).
  • SDP / SDR — Single Data Rate. Old IB speed (10 Gbps).
  • SHARP — Scalable Hierarchical Aggregation and Reduction Protocol. NVIDIA in-network compute (in-switch reduction for all_reduce). 1.5–2× speedup on collective ops.
  • slurmctld — Slurm controller daemon.
  • slurmd — Slurm node agent.
  • slurmdbd — Slurm accounting database daemon.
  • SM (1) — Subnet Manager (IB). Discovers topology, assigns LIDs, configures switches. See /docs/networking/subnet-manager.
  • SM (2) — Streaming Multiprocessor (NVIDIA GPU). Compute unit on the chip. H100 has 132 SMs.
  • SMP — Strategic Merge Patch. K8s default merge strategy for kubectl apply. Has the orphan-field-preservation gotcha. See /docs/kubernetes/argocd.
  • SMP (alt) — Symmetric Multi-Processing. Generic CPU term.
  • SOCKS — TCP-tunneling proxy protocol; commonly used for tenant SSH access.
  • SPANK — Slurm Plugin Architecture for Node and job Kontrol. Plugin hook system; pyxis is one.
  • SRIOV — Single Root I/O Virtualization. Hardware NIC partitioning into virtual functions.
  • SUNK — Slurm-on-Kubernetes (CoreWeave). Operator that runs Slurm controller/dbd/login as Pods. See /docs/slurm-sunk/intro.
  • SXM — NVIDIA proprietary GPU socket form factor (SXM4 = A100, SXM5 = H100/H200, SXM6 = B200).

T

  • TC — Traffic Class. Network priority queue (used for DCB/PFC).
  • TGI — Text Generation Inference. HuggingFace inference framework.
  • THP — Transparent Huge Pages. Linux feature; usually madvise mode for HPC. See /docs/kernel-tuning/transparent-hugepages.
  • TLB — Translation Lookaside Buffer. CPU page table cache; hugepages reduce TLB pressure.
  • TPU — Tensor Processing Unit (Google). Custom AI ASIC.
  • Triton — NVIDIA Triton Inference Server. General model-serving framework.

U

  • UCX — Unified Communication X. Communication framework underlying OpenMPI/RCCL multi-node.
  • UFM — Unified Fabric Manager (NVIDIA). Commercial appliance for IB fabric monitoring + provisioning.
  • UFW — Ubuntu Firewall. Wrapper around iptables; usually disabled on HPC nodes.
  • UPI — Ultra Path Interconnect (Intel). Cross-socket CPU bus (the modern replacement for QPI).
  • USD — Universal Scene Description (Pixar/Omniverse). 3D scene format.

V

  • VFIO — Virtual Function IO (Linux). Userspace driver framework; DPDK uses it.
  • vLLM — high-throughput LLM inference framework with paged attention KV cache.
  • VLAN — Virtual LAN. Ethernet L2 segmentation; less relevant in IB.
  • VRF — Virtual Routing and Forwarding. Per-tenant routing tables in IP networking.

W

  • Weka — distributed POSIX FS for HPC. See /docs/weka/architecture.
  • WUE — Water Usage Effectiveness. Liters of water per kWh of IT load. Liquid-cooled datacenters can have meaningful WUE.

X

  • XDR — Extreme Data Rate. IB speed (200 Gbps lane × 4 = 800 Gbps). Quantum-X800 / ConnectX-8 era.
  • Xid — NVIDIA driver structured error code. Critical ones: 13, 31, 43, 48, 63, 64, 74, 79. See /docs/operations/triage-decision-tree.

Y

  • YAML — the K8s format you've already typed today. Whitespace matters; tabs do not exist.

Z

  • Zero Trust — security model. Increasingly relevant as AI tenants run untrusted code.

See also