FAQ

Maintenance fees are mainly quoted based on the number of equipment, brand models, and service levels (response time, on-site availability, spare parts method). We provide various modes including annual full-machine maintenance, per-repair service, and single inspection. Free maintenance solutions and quotations can be provided based on your equipment list, with no hidden fees. Welcome to submit your equipment list online for a quotation.

Yes. We provide on-site inspection, repair and deployment services, covering key cities nationwide. Local engineers can arrive quickly; for remote areas, remote diagnosis can be handled first, and engineers can be arranged to travel if necessary. Specific coverage and on-site timings can be confirmed with customer service.

We provide 7×24-hour emergency response. Contracted maintenance customers enjoy priority response channels: phone/remote immediate access, remote intervention within 30 minutes for critical business faults, and on-site arrival time agreed by city level, minimizing business interruption to the maximum extent.

Repaired and replaced parts come with warranty. Original parts generally have a warranty of 3 months to 1 year (depending on part type), and repair workmanship also comes with warranty commitments. Same fault can be repaired free of charge within the warranty period. Specific warranty period will be clearly marked on the repair order.

Guaranteed. All second-hand equipment we sell undergoes strict testing (power-on self-test, stress testing, component health check), with condition and warranty status truthfully disclosed, and corresponding warranty services provided. Second-hand equipment offers high cost-effectiveness, suitable for budget-limited or testing/non-critical scenarios. Purchase with confidence.

Yes. We provide one-stop services from equipment racking, system installation, cluster deployment to data migration, supporting Oracle database migration, RAC setup, GPU computing platform deployment, etc. During migration, we focus on data security and business continuity, and can arrange maintenance window operations.

We have our own spare parts warehouse, stocked with a large number of original and used parts for NVIDIA and Oracle servers. For repair and replacement, we prioritize using original parts and truthfully inform the source of spare parts (brand new original/original used), ensuring quality traceability. Key spare parts come with warranty.

Yes. In addition to local on-site service, we serve customers in different cities through remote diagnosis, spare parts express delivery, on-site resident/assignment and other methods. Many long-term maintenance customers are distributed across the country. Through remote operation and maintenance platforms, efficient responses can be achieved, with remote service offering the same quality and price as local service.

B300 uses the Blackwell Ultra architecture with 288GB HBM3E per GPU, 8TB/s bandwidth and about 15 PFLOPS NVFP4, favoring large-model inference and long context. H200 uses Hopper with 141GB memory and about 1979 TFLOPS FP8, shares the H100 software stack, and suits mixed training and inference on a tighter budget. Choose B300 for inference-heavy, very large models; choose H200 for training-heavy work that reuses existing H100 experience.

Each B300 GPU has 288GB of HBM3E with 279GB usable and 8TB/s bandwidth. An 8-GPU DGX B300 system provides about 2.1TB of total GPU memory, roughly 144 PFLOPS FP4 and 72 PFLOPS FP8, nearly double the roughly 1.1TB available on an 8-GPU H200 system.

One H200 with 141GB can roughly serve a 70B-parameter model in FP16. An 8-GPU system with about 1.1TB total memory supports 400B-class models for multi-GPU inference or mid-scale training. With FP8/FP4 quantization, tensor parallelism and longer context, the practical model size changes and should be calculated from the model structure.

B300 draws about 1400W per GPU, with an 8-GPU system approaching 10-12kW at full load. Plan 1.3-1.5x power headroom, redundant supplies on separate PDUs and UPS backup. Cooling should be strong air or liquid, with hot and cold aisle separation and 20-25 degrees C intake. Assess power and cooling before deploying B300 in an older facility.

We supply from stock and build-to-order. Standard 8-GPU B300 and H200 servers are usually available quickly, while custom CPU, memory, storage and InfiniBand configurations are scheduled to order. Lead time depends on GPU supply and factory scheduling. For urgent projects, choose a standard in-stock configuration and expand later.

B300 is primarily shipped as an SXM module that needs an HGX baseboard with dedicated power and cooling, so it cannot be plugged into a standard PCIe server motherboard. SXM uses NVLink for high-bandwidth GPU-to-GPU communication and suits 8-GPU systems. PCIe cards install more flexibly in general servers but offer much lower interconnect bandwidth, so SXM is preferred for large-scale training and inference.

H200 and H100 share the Hopper architecture, and CUDA, kernel libraries and mainstream training frameworks are largely compatible, so application migration is cheap. However, H200 has 141GB memory and 700W power, so the baseboard, power and cooling must match H200 specifications; you cannot simply swap H100 cards in place. For expansion, buy H200 servers and interconnect them with the existing H100 cluster over InfiniBand.

A complete server quote generally includes GPU, CPU, memory, local storage, power supplies, fans and chassis. InfiniBand switches, cables, rack power distribution and software licenses are usually quoted separately. New servers carry manufacturer warranty, while refurbished modules are covered for 3 to 12 months. Confirm the warranty start date, coverage and spare-part response time in the contract.

Refurbished parts are viable if risk is controlled: verify GPU model, serial number and memory capacity, run memory diagnostics and extended stress tests, check driver logs for XID errors, observe full-load temperature and power, and confirm no repair traces on the module. Choose a supplier that provides test reports and at least 3 months warranty; use new servers for critical production.

B300 natively supports NVFP4 with about 15 PFLOPS dense compute, a major improvement over the previous generation. Using FP4 quantization in inference workloads where accuracy permits raises per-GPU throughput and reduces the number of cards required, lowering cost per token. Quantization does introduce accuracy loss, so evaluate it on your own workload data before production.

Inside the node, the eight H200 GPUs communicate over NVLink and need no external switch. For multi-node clusters, use InfiniBand such as 200G or 400G, or RoCE Ethernet, cabled rail-optimized with matching switches, optics and cables. Keep NIC and switch firmware versions aligned and validate real bandwidth and stability with an NCCL All-Reduce test before going live.

Before colocation, confirm available rack power and distribution type, UPS and redundancy, cooling capacity and intake temperature, rack weight and dimensions, network port rates and IP resources, and whether high-density equipment is allowed. An H200 system draws about 8-10kW and B300 more, so standard racks often cannot support it; high-density racks, strong air cooling or liquid cooling may be required. Confirm electricity and retrofit costs with the facility in advance.

Test environments can use refurbished machines and simplified networking; production requires redundant power, InfiniBand high-speed networking, redundant storage and maintenance coverage - stability first.

SPARC is Oracle's own RISC architecture, natively matched to Solaris and Oracle Database with strong reliability; x86 is a general-purpose architecture running Linux and Windows with a broader ecosystem.

There is no fixed number - it depends on CPU cores, memory and workload type. OLTP typically sustains thousands of concurrent connections; sizing should be based on a business-driven load test.

Solaris is the native and primary OS; some early versions could run a few Linux distributions, Windows is not supported, and mainstream workloads today run Solaris.

Yes. On-site hardware log collection and stress testing, with a hardware health assessment report.

Banks, insurance, securities, government and telecom legacy core transaction systems - mostly long-standing Oracle database environments.

Moderately risky: always back up first, work in a low-traffic window, and prepare a rollback plan.

Yes. Older SPARC models do not support new Solaris releases, and new Solaris versions have limited drivers for old hardware - versions must be matched at selection.

Four common categories: power faults (failed PSU or abnormal mains), board or CPU faults, loose or damaged memory (one of the most common causes), and RAID or expansion card faults. ZhongLianDa engineers can guide basic checks remotely (reseat memory, check power), with on-site testing for complex cases. We test first and quote later - no fix, no fee.

Yes, short-term and long-term rental for both AI servers and Oracle minicomputers.