Insights / Buying Guide

Do you pay for a B300 after a demo, or after about 18 hours?

Shell evidence chain for DGX-class B300 nodes — sampling, diagnostics, fabric, burn-in, and recovery

Quick Take

Quick Take: Accepting a B300 node means proving eight GPUs, their NVLink mesh, and the RoCE fabric all hold under sustained load before the invoice clears. The five-phase command chain below gives procurement the shell output, soak duration, and reset proof needed to reject a marginal node.

Once you sign the delivery receipt on a B300 server, any silent HBM3e ECC fault, NVLink lane degradation, or swapped NIC becomes your balance-sheet loss. Hold final milestone payment until the node passes every DCGM, nvbandwidth, and serial-entitlement gate below.

3-Line Fast RFQ Template (Ref: BQ-B300-ACC) — Copy and email to info@benquan.hk or open pre-tagged RFQ form →

Ref: BQ-B300-ACC (article-b300-acceptance-guide)
1) Delivered Server Model & GPU Count under Acceptance: 
2) Target Acceptance Gates (Serial/UDI check, DCGM L3, nvbandwidth): 
3) Delivery Location & Inspection Timeline: 

The Four Questions of B300 Acceptance

Quick Answer: Acceptance must answer what the machine is, what is installed, what is connected, and what evidence was captured. Skip any question and later benchmark scores lack a verifiable baseline for procurement disputes. A B300 node is typically an 8-GPU system.

A B300 node is typically an 8-GPU system. Map every phase to four questions: host identity (model, serial, firmware), installed inventory (GPUs, memory, storage, NICs), connectivity (NVLink, RoCE/IB, PCIe topology), and retained evidence (command output, logs, screenshots by phase).

OEM guidance puts quick health and pre-flight checks first. In NVSM environments, sudo nvsm stress-test --force is a useful pre-stress gate. Use DCGM for node diagnostics, nccl-tests for collectives, and Redfish ComputerSystem.Reset for recovery proof.

Phase 1: Raw Hardware Sampling

Quick Answer: Collect host model, serials, firmware, and nvidia-smi snapshots before any stress. Raw sampling is the reference you compare after burn-in, reseats, and part swaps — not a place to chase peak FLOPS. Establish asset, runtime, and network baselines before diagnostics: Retain host model/serial, GPU and disk counts, and NIC/port mapping.

Establish asset, runtime, and network baselines before diagnostics:

# System identity
dmidecode -t 1
dmidecode -t 17
dmidecode -t 39
hostnamectl
uname -a

# GPU inventory
nvidia-smi -L
nvidia-smi -q

# Bus and storage
lspci
lsblk

# Network and memory
ibstat
ibdev2netdev
free -h

Retain host model/serial, GPU and disk counts, and NIC/port mapping. Incomplete Phase 1 makes later bandwidth numbers non-auditable.

Phase 2: GPU and Intra-Node Diagnostics

Quick Answer: Verify GPU health, NVLink topology, and intra-node bandwidth together. A short demo with broken NVLink, missing peers, or rising ECC still fails Phase 2 even if one matrix multiply looks fine. dcgmi diag -r 4 covers memory, interconnect, and general health.

# Comprehensive GPU diagnostics
dcgmi diag -r 4

# Topology and NVLink status
nvidia-smi topo -m
nvidia-smi nvlink --status

# Intra-node bandwidth
nvbandwidth
p2pBandwidthLatencyTest

dcgmi diag -r 4 covers memory, interconnect, and general health. Confirm all eight GPUs and expected NVLink mesh before you trust multi-GPU training claims.

Phase 3: RoCE and NCCL Cross-Node

Quick Answer: Phase 3 moves testing off the node: confirm every port reports active with ibstat and ethtool, measure point-to-point throughput with ib_write_bw and ib_read_bw, then run all_reduce_perf across hosts up to 4 GB payloads. Prove the lossless queue drops nothing under sustained pressure — a console screenshot showing PFC and ECN enabled proves little.

# Link and RoCE configuration
ibstat
ibdev2netdev
ethtool <interface>
mlxconfig query

# Point-to-point throughput and latency
ib_write_bw -a
ib_read_bw -a
ib_send_lat -R

# NCCL collectives (example)
./build/all_reduce_perf -b 8 -e 4G -f 2 -g 8

For RoCE, validate PFC, ECN, and DCQCN together — switch features enabled on paper do not guarantee a lossless queue in practice.

Phase 4: Long-Duration Stability Burn-In

Quick Answer: Soak the node for roughly 18 hours — gpu_burn with a 64,800-second timer — while nvidia-smi dmon logs power, clocks, and temperature every minute and journalctl is grepped for XID events. Faults that surface only after hours of heat, such as ECC growth or link flaps, are exactly what a ten-minute demo hides.

# 18-hour GPU burn-in (example)
gpu_burn -tc 64800

# Continuous monitoring (60s intervals)
nvidia-smi dmon -s pucvmet -d 60

# XID monitoring
journalctl -k | grep -i xid

This phase answers “does it hold when hot,” not “can it boot.” Watch temperature peaks, fan curves, correctable ECC, and XID entries across the soak window.

Phase 5: Recovery Path Validation

Quick Answer: Document and test BMC/Redfish reset before go-live. An unresponsive node without a proven recovery path is operational debt, not accepted inventory ready for the cluster schedule. A verified recovery chain beats gambling on a physical power cycle during a production incident.

# Redfish system reset (example)
curl -X POST https://<bmc-ip>/redfish/v1/Systems/System.Embedded.1/Actions/ComputerSystem.Reset \
  -d '{"ResetType": "GracefulShutdown"}' \
  -H "Content-Type: application/json"

A verified recovery chain beats gambling on a physical power cycle during a production incident.

Organizing the Acceptance Script

Quick Answer: Map scripts to five roles — sampling, node diagnostics, fabric, burn-in, and recovery — with clear owners. Auditable evidence matters as much as pass/fail when finance and vendors dispute a return. dcgmi diag, nvbandwidth, topo, P2P tests. Role — Commands — Purpose Raw Sampling — dmidecode, nvidia-smi, lspci, ibstat, lsblk — Asset and baseline capture Node Diagnostics —.

Role Commands Purpose
Raw Sampling dmidecode, nvidia-smi, lspci, ibstat, lsblk Asset and baseline capture
Node Diagnostics dcgmi diag, nvbandwidth, topo, P2P tests GPU health and NVLink
Fabric ib_*, mlxconfig, nccl-tests Cross-node collectives
Burn-In gpu_burn, dmon, XID logs Sustained stability
Recovery Redfish ComputerSystem.Reset Documented reset path

Conclusion

Quick Answer: B300 acceptance is an evidence chain: baseline, GPU and fabric proofs, long burn-in, then recovery. A node that passes all five phases with retained logs is ready for production AI workloads. Skip any layer and secondary or new fleets become silent scrap risk.

Skip any layer and secondary or new fleets become silent scrap risk. BENQUAN GLOBAL LIMITED supplies B300 with configuration support for Hong Kong and APAC — including acceptance-oriented procurement guidance.

How to Run the Five Phases in Order

Quick Answer: Run the phases in sequence and keep the output of each before starting the next: baseline first, then GPU and NVLink diagnostics, cross-node RoCE and NCCL, an 18-hour burn-in, and finally a Redfish reset test. Do not start burn-in until Phases 1 to 3 are clean.

  1. Step 1 - Raw sampling: capture dmidecode, nvidia-smi, lspci, lsblk and ibstat output as the baseline.
  2. Step 2 - GPU and node diagnostics: run dcgmi diag -r 4, nvidia-smi topo -m and nvbandwidth; confirm all eight GPUs and the NVLink mesh.
  3. Step 3 - RoCE and NCCL: confirm active ports, run ib_write_bw and all_reduce_perf across hosts, and check PFC, ECN and DCQCN behaviour.
  4. Step 4 - Burn-in: run gpu_burn -tc 64800 (18 hours) with nvidia-smi dmon logging and XID monitoring.
  5. Step 5 - Recovery: test a Redfish ComputerSystem.Reset and record the result before sign-off.

Frequently Asked Questions

Quick Answer: Plan multi-hour diagnostics plus about 18 hours of burn-in, keep Phase 1–5 evidence, and treat missing recovery tests as a fail. Contact BENQUAN GLOBAL LIMITED for B300 RFQ and acceptance support worldwide. About 18 hours of full-load burn-in is a practical minimum.

B300 acceptance is a five-phase protocol: raw hardware sampling, GPU and intra-node diagnostics, RoCE and NCCL cross-node checks, about 18 hours of stability burn-in, and recovery-path validation. Skip a layer and later fleets become silent scrap risk.

About 18 hours of full-load burn-in is a practical minimum. Short demos miss thermal and ECC issues that appear only under sustained heat.

GPU health, NVLink topology, and intra-node bandwidth together. Broken links or ECC growth fail the phase even if a short demo looks green.

Key Takeaway

Key Takeaway: Do not sign acceptance until every phase carries retained logs: baseline serials and firmware, dcgmi diagnostics showing a complete NVLink mesh, NCCL all-reduce across nodes, roughly 18 hours of clean soak, and one successful Redfish reset. Missing evidence turns a warranty claim into an argument.

Not sure your BOM is optimal?

If you are unsure whether your BOM is optimal, click below for a free compatibility and delivery risk assessment. Fast global BOM validation and risk assessment.

Free Compatibility & Delivery Risk Assessment

Or call +852-21349834 · info@benquan.hk

Contents