Turn “eight GPUs” into a specific requirement
Prepare one requirements sheet for each quotation: inference, fine-tuning, training or rendering; exact application or model version; weight precision; input and output length; concurrent requests; response-time target; data size; quantity and delivery location. These are procurement checklist items, not certified capabilities of a particular server. If the workload is undecided, start with representative tasks for a technical assessment. An eight-GPU label is not a guarantee that every model or workload will run.
Confirm the exact GPU and physical form factor
Do not treat different form factors in one GPU family as interchangeable. NVIDIA lists H100 SXM and the PCIe-form-factor H100 NVL with different power and interconnect specifications. This example explains why the distinction matters; it does not certify either variant for every server on this website. Record the exact GPU part number, memory per GPU, quantity, slot width and server module. Confirm the vendor support list, power cables and cooling requirements. Eight physical positions do not establish compatibility with any eight GPUs.
Check direct connection, PCIe switches and NVLink separately
H3C lists separate eight-GPU Switch, four-GPU direct, eight-GPU direct and sixteen-GPU Switch modules for the R5300 G6. Record the actual module and slot allocation; do not combine maxima from different columns into one configuration. PCIe-switch topology and GPU NVLink support are separate checks. Request the CPU, GPU and network-adapter connection layout. Validate multi-GPU communication with the intended software and workload rather than assuming performance from a GPU-count or Switch label.
Aggregate memory is not automatically one request’s memory pool
Multi-GPU deployment needs an explicit software strategy. The vLLM documentation distinguishes tensor and pipeline parallelism and describes startup estimates for KV-cache capacity and concurrency at a specified request length. Record the actual parallel configuration and inspect weights, KV cache and runtime memory separately. Test realistic input/output lengths, concurrency and precision, keeping per-GPU utilization and logs. Neither aggregate memory alone nor a blanket single-GPU limit is a sufficient deployment assessment. This guide makes no unmeasured throughput or concurrency commitment.
Include facility and redundancy requirements in the quotation
Check usable rack depth, rails, rear cable clearance, airflow, electrical circuits, PDU, plugs and available cooling. Request a system-power assessment, redundancy policy and the exact power-supply configuration being delivered. H3C notes that the recommended R5300 G6 redundancy strategy can vary with power-supply output and refers buyers to the technical white paper or support. Four installed power supplies should not be interpreted as a universal redundancy guarantee. These are procurement checks; qualified personnel should confirm electrical and safety conditions.
Accept the system against model-specific documents and a complete BOM
The related-product links lead to the H3C R5300 G6 and Lenovo WA5480 G3 pages, where model-specific FAQs and documents can be checked. They are product references, not unconditional configuration recommendations. Do not transfer one model’s CPU, storage, GPU or power options to another. Suggested acceptance records include a complete bill of materials and serial numbers, detected GPU quantity, firmware and driver versions, parallel configuration, workload test conditions, logs, exception handling and service responsibilities. Recheck when documentation changes. Price, availability, delivery and service scope are governed by the written quotation and contract.
Frequently asked questions
Select a question to expand its answer.
Does an eight-GPU server necessarily include NVLink?
No. Check the exact GPU, interconnect components and server topology separately. A PCIe-switch configuration does not by itself establish NVLink support.
Why inspect memory on each GPU when all eight are installed?
Per-GPU use depends on how the software distributes weights and cache. Evaluate the actual parallel strategy, request length, concurrency and runtime logs, not just aggregate memory.
Can I install fewer GPUs now and expand to eight later?
Include the expansion target in the pre-purchase BOM review. Confirm modules, slots, power, cables, cooling and firmware. An eight-GPU product label alone does not prove that the current configuration contains everything needed for expansion.
What information should I provide for a quotation?
Provide the workload or model version, precision, input/output length, concurrency target, preferred GPU, data size, quantity and facility power/cooling conditions. Mark unknown items as unconfirmed rather than inventing values.
Sources & scope
Guides do not replace the exact model documentation, verified bill of materials or a written quotation.
- Lenovo WenTian: WA5480 G3 manufacturer brochure (pages 1–2, ©2024; original file hash checked)www.infr-tech.com · Checked 2026-11-01
- H3C: R5300 G6 specifications and GPU compute moduleswww.h3c.com · Checked 2026-11-01
- NVIDIA: H100 specifications and form factorswww.nvidia.com · Checked 2026-11-01
- vLLM: Parallelism and Scalingdocs.vllm.ai · Checked 2026-11-01
- vLLM: KV-cache configurationdocs.vllm.ai · Checked 2026-11-01