Computing–network convergence simulation platform

A simulation test platform for AI large-model training and computing cluster networks, covering InfiniBand, ROCEv2 and lossless Ethernet.

One-line positioning

A simulation test platform for AI large-model training and computing cluster networks, covering InfiniBand, ROCEv2 and lossless Ethernet.

Product overview

XproHPC serves AI large-model training, high-performance computing and computing cluster networks. It combines hardware simulation, protocol analysis and collective communication algorithms to test mainstream computing network technologies such as InfiniBand, ROCEv2 and lossless Ethernet, from single nodes to large-scale clusters.

Core capabilities

Computing network testing

  • GPU NCCL collective communication simulation
  • RDMA protocol testing
  • Lossless Ethernet (PFC / ECN) validation
  • User-defined communication patterns, message sizes, traffic models and link impairment scenarios to construct abnormal network conditions, evaluate the stability of switches, NICs and computing clusters, and identify network performance bottlenecks

Collective communication simulation

  • Standard NCCL primitives: AllReduce, Broadcast, Reduce, Gather, Scatter, AllGather, AllToAll and Ring-Reduce
  • Topology algorithms such as Ring, Tree and Double Binary Tree
  • Configurable message sizes, communication intervals and burst periods

Traffic modeling

  • Configurable message sizes from 1B to 16MB
  • Communication data types: character, integer and floating point
  • Configurable GPU training wait time to simulate burst traffic
  • Continuous and custom burst traffic (burst packet length, interval and rate threshold)

Delay measurement and synchronization

  • One-way forwarding delay measurement precision of 100 nanoseconds, with measurement across ports and chassis
  • IEEE 1588v2 PTP-based nanosecond-level time synchronization, with cross-board / cross-cabinet precision ≤100ns
  • Nanosecond-level timestamp capture and marking
  • Up to 100GB of high-speed data buffering, line-rate packet capture and continuous storage, PCAP export and offline analysis

Computing-network convergence capabilities

  • Computing cluster topology simulation: three-tier Clos architecture (Leaf-Spine-SuperSpine), supporting star, dual-ring and binary-tree topologies, with visual location of failed nodes, abnormal traffic and parameter errors
  • Cross-data-center testing: WAN RDMA lossless transmission testing and cross-domain NCCL collective communication optimization validation
  • SimAI integration: a closed loop from “simulation topology generation—traffic construction—test execution—result analysis”, testing both computing and communication efficiency, with simulation error <2%
  • High-volume packet capture and analysis: 400G-bandwidth out-of-band statistics, millisecond burst traffic statistics, protocol parsing, session tracking and anomaly detection
  • Dedicated ECN / PFC testing: per-port counts of ECN-marked packets, transmitted and received CNP control traffic, PFC Pause frames and pause duration, and out-of-order packets

Key performance specifications

Item Specification
Maximum simulated nodes per instrument 24 computing servers
Maximum communication bandwidth per instrument 4.8 Tbps
Port density 24×200GE or 12×400GE per chassis
Expansion Cross-chassis cascading with linear bandwidth scaling
Delay measurement precision 100-nanosecond-level one-way forwarding delay
Time synchronization precision ≤100ns across boards / cabinets
Supported protocols InfiniBand / ROCEv2 / PFC / ECN / NCCL
Collective communication modes AllReduce / Broadcast / Reduce / Gather / Scatter / AllGather / AllToAll / Ring-Reduce
Message size range 1B – 16MB
Traffic models Continuous, burst and incast traffic

Deployment requirements

  • Deployment: X86 server + NVIDIA BlueField-3 DPU, as integrated hardware and software (virtualized and cloud deployment are not supported)
  • CPU: ≥20 cores (≥48 threads recommended for performance tests)
  • Memory: minimum 4GB, ≥128GB recommended
  • Hardware interfaces: NVIDIA BlueField-3 DPU ports, with mixed 100GE / 200GE / 400GE interfaces
  • Driver: DPDK driver binding
  • Licensing: USB KEY activation / software lock, with License migration

Ordering models

Model Port configuration
XproHPC-4.8T 12×400GE
XproHPC-3.2T 8×400GE
XproHPC-2.4T 6×400GE
XproHPC-1.6T 4×400GE
XproHPC-800G 4×200GE or 2×400GE
XproHPC-400G 2×200GE or 4×100G

Typical test scenarios

  • InfiniBand and ROCEv2 deployment validation and performance tuning in intelligent computing / supercomputing centers
  • Collective communication efficiency testing for AI large-model training clusters
  • Lossless Ethernet and ROCEv2 compatibility validation for switch vendors
  • Cross-data-center computing expansion and collaborative training testing for operators
  • Computing network research and protocol validation for research institutes
  • Communication reliability assurance testing for distributed AI large-model training