Hardware sources and compatibility scope
Prepared on September 21, 2026. The rules correspond to the editions consulted on that date. Links lead to the original Lenovo Press pages; cited page numbers identify the consulted edition and may differ in later revisions.
Lenovo ThinkEdge SE455 V3
Product Guide: LP1724. Main sections: pages 18–21, 37–41.
- Initial selection is limited to 1–2 GPUs to reserve network expansion. The appropriate risers are required.
- L40S: CPU ≤130 W, 1U heatsink, ambient temperature ≤40 °C, no acoustic mode. A full-length kit and GPU cabling are required; check the 1800 W power supply.
Lenovo ThinkSystem SR680a V4
Product Guide: LP2264. Main sections: pages 12, 15–16.
- Fixed HGX B300 complex with 8 GPUs; H100 or H200 cannot be selected in this entry.
- 64 and 96 GB: 32 DIMMs required. 128 GB: 16 or 32 DIMMs. All must be identical.
Lenovo ThinkSystem SR680a V3
Product Guide: LP1909. Main sections: pages 12, 17–19.
- The guide identifies this server as a withdrawn product.
- 8 GPUs from the same HGX complex. The Gaudi 3 variant is not included in this initial version.
- H100/H200 with 64 GB DIMMs requires 32 modules.
Lenovo ThinkSystem SR780a V3
Product Guide: LP1980. Main sections: pages 10–15.
- The GPU board and server base code must match.
- Liquid cooling installation is required. Front network adapters: groups of 4 or 8 identical adapters; the displayed figure is per port.
Lenovo ThinkSystem SR685a V3
Product Guide: LP1910. Main sections: pages 10–15.
- 24 DIMMs required: 12 per processor, all with the same part number.
- The fifth-generation CPU enables only the C0CL 128 GB DIMM in the consulted guide. A conservative speed of 6000 MT/s follows its memory rules.
Lenovo ThinkSystem SR675 V3
Product Guide: LP1611. Main sections: pages 23–28, 44–49.
- This version offers bases with 4 or 8 PCIe GPUs and two CPUs. Network slots depend on the base, risers, and storage.
- H200 NVL is excluded from this version until its specific memory and cooling rules are incorporated.
Lenovo ThinkSystem SR650a V4
Product Guide: LP2128. Main sections: pages 16–18, 24–25, 68–71.
- Two CPUs and identical DIMMs distributed equally across both sockets. No MRDIMMs or mirrored memory in this version.
- L4 requires a single-GPU base/riser; L40S requires a double-GPU base/riser. Validate the thermal rules for the selected CPU.
Lenovo NVIDIA GB300 NVL72 Rack Scale AI
Product Guide: LP2357. Main sections: pages 5, 18–19, 22–25.
- Calculation unit: a complete rack, 18 trays, 36 CPUs, and 72 GPUs. Fixed RAM: 960 GB per tray.
- The NVLink domain does not extend across racks. A CDU and power design are required; the guide states 135 kW TDP and up to 155 kW peak.
Validation scope
The catalog is a partial selection of explicitly listed CPUs, memory, and GPUs. It does not combine different CPUs or DIMMs with different part numbers. It filters RAM by CPU generation, limits quantities, and blocks L40S with CPUs above 130 W in the SE455. HGX/OAM accelerators and the rack have fixed quantities.
Not every matrix of risers, slots, power supplies, backplanes, cables, temperature, firmware, or commercial availability has been encoded. Selecting components does not constitute a certified configuration for purchase. PCIe networking is shown as documented options; installation and quantity depend on slots, risers, and bases. A complete BOM is not calculated.
SR675: symmetric RAM selection of 4, 8, or 12 DIMMs per CPU. SR650a: two CPUs and RDIMMs only; MRDIMMs, 600 W GPUs, and their special kits are omitted. SE455: configurations with 3 to 6 L4 GPUs are omitted. SR680a V3: Gaudi 3 is omitted. GB300: 36 × 480 GB RAM = 17,280 GB per rack; the inconsistent mention of 8 ×128 GB in its guide is not interpreted as usable capacity.
Models and calculation — v0.2
Qwen models retain the official configurations linked in the calculator catalog. Q heads, vocabulary, and a BF16 embedding/output budget derived from those dimensions are added.
Methodology v0.1 is superseded by METODOLOGIA.html: complete formulas, assumptions, and calculation limits. The engine now evaluates TP/PP, replicas, and the budget per GPU; it does not combine all VRAM as if it were shared memory.
The physical Lenovo options have not changed in this revision.