FAQ

A security list applies to an entire subnet; an NSG applies to an individual VNIC. They are independent firewalls — if you use NSGs, changing only the security list has no effect. Ports must be allowed on both sides.

Normal, compliant use does not trigger auto-shutdown; instances get disabled only when quotas are exceeded or risk control fires. There is no 'anti-shutdown script' — the most effective prevention is keeping resource usage below the free limits.

A1-Flex ARM is stronger and suits web services and applications; E2-Micro x86 offers better compatibility for software that only runs on x86. Prefer A1 under the free allowance, and switch to E2-Micro if you hit software compatibility issues.

Stop: the machine powers off, disks are retained and continue consuming quota. Terminate: the VM is destroyed completely. Note that stopping an ARM instance releases its ephemeral public IP, and disks consume quota whether the machine is on or off until deleted.

There is no strict instance-count limit; the constraint is total OCPU and memory quota. A1-Flex tops out at 4 cores and 24 GB — you can split that into several small machines, as long as the total stays within the free allowance.

When the $300 trial expires, paid resources stop being billed while Always-Free resources remain untouched. However, machines exceeding the free spec are stopped after expiry — remember to downsize your configurations in time.

Use force-terminate for stuck instances. Note that deleting the VM is not the whole cleanup: you must also manually delete attached disks, reserved public IPs, object storage buckets and other related resources — leftovers keep generating charges.

The free tier blocks outbound TCP port 25 entirely, so self-hosted mail servers cannot send. Paid accounts can request port 25 unblocking via a support ticket. Free users should use a third-party email provider instead of building their own SMTP service.

The default public IP is ephemeral and changes on restart or stop. To fix it, create a Reserved Public IP in the console and bind it to the instance's VNIC — a reserved IP stays unchanged until you manually release it.

Common causes: incorrect SSH key file permissions, wrong PuTTY ppk key format, or the OCI security list not allowing port 22. Free-tier instances may also be blocked by risk control. If SSH fails, use the serial console to log in and handle the issue.

Yes — remotely or on-site, including Solaris system log review and basic fault localization.

It depends on the component: power supplies and fans are hot-swappable with no downtime; core parts such as motherboards and memory require shutdown — typically tens of minutes to a few hours.

We recommend quarterly preventive maintenance: update firmware/BIOS to stable versions (improving compatibility and security), clean chassis dust (poor cooling accelerates aging), check and tighten loose data and power cables, replace power-supply capacitors older than 3 years of runtime, and verify backup restorability. ZhongLianDa maintenance contracts include preventive maintenance: we agree on a maintenance window in advance, take full backups before operating, and deliver a report afterwards.

Hardware health status, temperature and power draw, error logs, firmware versions, risk items, optimization recommendations and an inspection summary.

We operate our own spare-parts warehouse stocking large quantities of NVIDIA and Oracle server OEM parts and genuine pull-outs. Repairs prioritize OEM parts, and we disclose part origin honestly (new OEM / genuine pull-out) so quality is traceable — key parts come with warranty.

Hardware alerts, power/fan status, memory, drive health, temperature, firmware versions, system logs, and RAC status.

We provide 7×24 emergency response. Contract maintenance customers get a priority channel: immediate phone/remote access, remote engagement within 30 minutes for critical business faults, and city-tier-based on-site arrival commitments — minimizing business downtime.

Routine inspection covers hardware, software and environment. Hardware: IPMI/iLO remote checks of CPU temperature, memory usage, drive health, power redundancy and fan speed. Software: abnormal system logs, RAID controller status, network port traffic, time synchronization. Environment: room temperature and humidity, UPS battery state, fire alarms. ZhongLianDa contract customers receive on-site inspection at least monthly, plus daily automated inspection reports from our remote monitoring platform with instant alerts on anomalies.

Customer reports by phone or ticket → remote engineer triages and locates the fault → on-site need assessed → field handling → fault report submitted after repair.

Training: high-speed GPU interconnect (NVLink/IB), large memory, high power draw. Inference: lower interconnect requirements, can lean on CPU compute, lower power, and supports high-density deployment.

A100: mature, for large-model training/fine-tuning and existing projects. H100: high compute for very large model training, with clear FP8 advantages. B300: the mainstream choice in China, balancing training and inference with strong value. B100: next-generation high-end card for very large compute clusters.

A 7B small model: one 8-GPU machine suffices. A 34B model: typically 2–4 machines. 70B and larger: at least 4 machines — the bigger the model, the larger the required cluster.

Not necessarily. With model quantization, B10, A10 or even T4 cards can serve inference — choose by actual need instead of defaulting to costly cards like H100.

In large-model training, NVIDIA H100 holds a clear advantage: the Transformer Engine speeds training 2–3x, FP8 precision accelerates further while maintaining accuracy, and memory bandwidth rises from 2 TB/s to 3.35 TB/s. Training a 70B-parameter model, an H100 8-GPU cluster can cut the cycle from months to weeks; in inference, H100 delivers roughly 2x throughput. ZhongLianDa keeps H100 in stock and can set up a test environment so you can verify the difference yourself.

Both exist: branded servers carry whole-system warranty; separately purchased GPU boards carry their own board-level warranty. Second-hand and third-party maintenance typically covers the GPU board individually.

GPU VRAM errors fall into two kinds: correctable ECC errors (usually transient interference — logged and ignorable) and uncorrectable ECC errors (possible VRAM chip hardware faults or GPU core issues). The latter requires repair. ZhongLianDa runs a professional GPU repair team with BGA rework stations and VRAM chip stock, fixing common issues such as cold solder joints and failed memory chips. Repaired units pass a 24-hour stress test before return and carry a 3-month warranty.

Check fans, airflow and machine-room intake temperature; clean dust; inspect the GPU cooler module; adjust workload and optimize room air conditioning.

Five common causes: data-loading bottleneck (CPU can't feed the GPUs), multi-GPU communication bottleneck (misconfigured InfiniBand), insufficient VRAM causing frequent paging, batch size too small, and low parallel efficiency of the model itself. ZhongLianDa offers a free GPU utilization diagnosis — we collect runtime data with NVIDIA DCGM, locate the bottleneck and deliver an optimization plan. We also help tune cluster scheduler parameters and data pipelines, typically raising GPU utilization from 30–50% to above 80%.

Common causes: unstable power delivery, poor PCIe contact, BIOS settings, abnormal GPU firmware; in a few cases the card itself is damaged. Start by reseating the card and resetting BIOS defaults.

GPU not detected by nvidia-smi, card dropouts, random crashes during training, thermal throttling, ECC errors, VRAM errors, abnormal compute performance, black screens.