AI / Datacenter

Kernel-bypass networking model: queue pairs, memory regions, and work requests move data between registered buffers with minimal CPU. Com…

RDMA over UDP/IPv4/IPv6 on Ethernet: InfiniBand BTH follows the UDP header (port 4791). Routable across L3 with ECN/PFC/DCQCN for congest…

Lossless switched fabric with its own L2/L3 (LID/GID), subnet manager, and verbs transport (RC/UD/…). Credits, VLs, and partition keys pr…

Ubiquitous LAN frame: 6-byte MACs, 2-byte EtherType, payload, and 32-bit FCS. Preamble/SFD precede the frame on the wire; 802.1Q inserts …

Reliable, ordered byte-stream transport with ports, sequence/ack numbers, windowing, and congestion control. Three-way handshake and four…

Platform-independent industrial information model with secure channels, sessions, subscriptions, and monitored items. Binary UA-TCP and H…

RDMA over Converged Ethernet version 1 encapsulates InfiniBand verbs in Ethernet frames without IP/UDP headers. It is an L2-only design l…

RDMA over TCP/IP using MPA framing with DDP and RDMAP layers. It traverses lossy networks better than RoCE but needs different NIC offloa…

Lossless and ECN tools for datacenter Ethernet: Priority Flow Control pauses individual priorities, while Explicit Congestion Notificatio…

Data Center Quantized Congestion Notification — RoCE congestion control that uses ECN marks and Congestion Notification Packets to rate-l…

NIC DMA path that reads and writes GPU memory directly, bypassing host bounce buffers for RDMA verbs. Requires peer-to-peer BAR mapping, …

Linux buffer-sharing primitive for zero-copy exchange across drivers such as GPU, NIC, and media. Exporters and importers negotiate acces…

NVIDIA Collective Communications Library implementing allreduce, broadcast, and related GPU collectives over NVLink, PCIe, and RDMA. Topo…

AMD ROCm Collective Communications Library — the ROCm counterpart to NCCL for GPU collectives. It uses Infinity Fabric, PCIe, and RDMA pa…

Message Passing Interface — portable parallel messaging with point-to-point and collective operations across ranks. Implementations map t…

Scalable Hierarchical Aggregation and Reduction Protocol — Mellanox/NVIDIA in-network aggregation for MPI/NCCL collectives on InfiniBand …

RPC framework on HTTP/2 with protobuf stubs, streaming RPCs, deadlines, and metadata. Widely used as the control plane for ML serving, mi…

Multiplexed binary framing over a single TLS/TCP connection with streams, HPACK headers, and flow-control windows. It underpins gRPC and …

HTTP/gRPC model-inference APIs popularized by Triton/KServe V2: readiness/liveness, infer, and model-repository control. Tensor contents,…

gRPC protocol for bulk Apache Arrow record-batch transfer with DoGet/DoPut and FlightDescriptors. Columnar data moves with minimal copies…

NVIDIA high-bandwidth GPU interconnect for peer-to-peer and coherent-ish memory traffic inside a node or NVSwitch domain. Generations dif…

Switch fabric chip that stitches many NVLink ports into all-to-all GPU topologies in DGX/HGX-class systems. Enables full-bandwidth collec…

Compute Express Link — a cache-coherent interconnect riding PCIe PHY for memory expanders and accelerators. .io, .cache, and .mem protoco…

Universal Chiplet Interconnect Express — a die-to-die standard with streaming and protocol mappings (including PCIe/CXL). Advanced packag…

PCI Express — packetized serial I/O with lanes, LTSSM link training, and TLP/DLLP layers. Gen5/Gen6 rates and retimers define reach; ACS …

NVMe over Fabrics extends NVMe queues across RDMA, TCP, or Fibre Channel. Capsules map submission/completion queue entries over the fabri…

GPUDirect Storage (GDS) provides a direct DMA path between NVMe and GPU memory, skipping host bounce buffers. Filesystem and driver suppo…

User Datagram Protocol — connectionless transport with ports and an optional checksum. It is the basis for DNS, RTP, QUIC, DHCP, and RoCE…

NVM Express — PCIe register and queue interface for SSDs with submission/completion queues, PRP/SGL data pointers, and admin commands. Pa…

Protocol Buffers — language-neutral binary serialization with tagged fields and disciplined schema evolution. Default encoding for gRPC r…