What "NVIDIA-Certified" actually means

NVIDIA-Certified Systems is a validation programme: NVIDIA and the server OEM jointly test a specific server-plus-GPU-plus-networking configuration and certify that it delivers the expected performance, manageability and security for AI and data-analytics workloads. It is a badge on a tested configuration, not a blanket claim about a chassis. The practical value to a buyer is assurance: a certified configuration is a known-good pairing of server, GPU and firmware, so you are not the one discovering a thermal or PCIe-lane limitation in production.

Separately, systems built around the NVIDIA HGX baseboard (the 4-GPU or 8-GPU SXM board with NVLink/NVSwitch) are a distinct, higher tier — these are the machines for training models that must span multiple GPUs.

Reading a GPU-server spec sheet

Six fields decide whether a server fits the job:

• Form factor (rack U). Density and cooling headroom. A 2U accelerator server (R760xa, DL380a) holds up to ~4 double-width GPUs; a 4U (Supermicro 421GE) up to ~10 PCIe GPUs; an 8U (Supermicro 821GE) carries the full 8-GPU SXM baseboard.
• GPU capacity — double-width vs single-width. Data-centre GPUs are double-width (dual-slot) cards. "4 double-wide or 8 single-wide" means four full-power accelerators, or eight thinner cards. This is the number that matters most.
• PCIe vs SXM. PCIe = cards in slots, air-coolable, flexible. SXM = socketed modules on an HGX baseboard with NVLink — mandatory for large multi-GPU training, but needs a purpose-built (often liquid-assisted) chassis.
• CPU platform. Current generation is Intel Xeon Scalable 4th/5th Gen or AMD EPYC 9004/9005, all with DDR5 and PCIe Gen5. PCIe Gen5 doubles per-lane bandwidth to the GPU over Gen4 — it matters for GPU-dense boxes.
• Memory and drives. DDR5 capacity for the CPU side; NVMe Gen5 bays for feeding data to the GPUs.
• Power and cooling. GPU servers are high-wattage. An 8-GPU SXM node is a 10 kW-class machine; even a 4-GPU PCIe 2U wants multi-kilowatt redundant Titanium PSUs and serious airflow.

The mainstream models compared

Model	Form	Sockets / CPU	Max double-width GPUs	GPU topology	Notable
Dell PowerEdge R760xa	2U	2× Intel Xeon 4th/5th Gen	4 (or 12 single-width)	PCIe Gen5	GPU bays moved to chassis front for cold-air-first cooling; up to 2,800 W Titanium PSU
HPE ProLiant DL380a Gen11	2U	2× Intel Xeon 4th/5th Gen	4 (or 8 single-wide)	PCIe Gen5	'a' = accelerator-optimised; supports H100, H200 NVL, L40S; up to 3 TB DDR5
Lenovo ThinkSystem SR675 V3	3U	2× AMD EPYC 9004/9005	8	PCIe or HGX 4-GPU SXM5	One model does either 8× PCIe (H200/L40S) or the 4-GPU NVLink SXM baseboard
Supermicro SYS-421GE-TNRT	4U	2× Intel Xeon 4th/5th Gen	10	PCIe Gen5 (dual-root)	High-count PCIe workhorse; H100/A100/L40S; 4× 2,700 W Titanium
Supermicro SYS-821GE-TNHR	8U	2× Intel Xeon 4th/5th Gen	8 (SXM)	HGX 8-GPU SXM5 + NVSwitch	Flagship NVLink node: HGX H100 80GB or HGX H200 141GB; 6× 3,000 W PSUs

Choosing between them

• A mainstream on-prem AI/pro-viz build, standard rack, air cooling → Dell R760xa or HPE DL380a Gen11. Two-socket, 2U, up to four double-width GPUs (H100 PCIe, H200 NVL, L40S). The pragmatic majority of enterprise AI purchases land here.
• More PCIe GPUs in one box, still air-cooled → Lenovo SR675 V3 (8× PCIe) or Supermicro SYS-421GE-TNRT (up to 10× PCIe). Good for inference farms and mixed AI+graphics at scale.
• Multi-GPU training that must pool GPU memory over NVLink → the HGX SXM machines: Lenovo SR675 V3 in its 4-GPU SXM form, or the Supermicro SYS-821GE-TNHR 8-GPU node. These are the training-class systems and carry training-class power, cooling and price.

A note for the refurbished/new buyer

All five models above are current-generation (2023 platforms: Intel Xeon 4th/5th Gen or AMD EPYC 9004/9005, DDR5, PCIe Gen5) and are the safe modern floor for an AI build. Their prior-generation siblings — Dell R750xa, HPE DL380 Gen10 Plus, Lenovo SR670 V2, Supermicro X12 GPU systems — are the cheaper refurbished tier but sit on PCIe Gen4 and DDR4, which throttles GPU bandwidth and caps the newest accelerators. The 8-GPU SXM class (Supermicro 821GE and equivalents) is documented here for completeness, but a full HGX node is rarely a mainstream refurbished-channel item — treat it as a specify-and-source, build-to-order machine rather than an off-the-shelf refurb.

Always confirm the exact GPU support, PSU rating and cooling class against the model's current OEM QuickSpecs / spec sheet before quoting — GPU compatibility lists are revised as new accelerators (H200, B200) are qualified.