NVIDIA-Certified Systems: what the label covers and how to read the list
Eurokommerz, Vienna, since 2006: Private AI/ML · IT Managed Services · Enterprise Training · AI Hardware & Software
- An NVIDIA-Certified System is a server, workstation or edge system from a system maker that has passed NVIDIA's certification programme with the NVIDIA GPUs named in its row of NVIDIA's list
- NVIDIA's partners run the tests in a standardised setup and NVIDIA engineers evaluate the systems for performance, functionality, scalability and security; NVIDIA names no test tool and publishes no pass thresholds
- A row of the list names the maker, the model and the GPUs; reference configurations such as PCIe Optimized 2-8-5-200 add the tested network devices, and a footnote can state that GPU power is capped
- The NVIDIA AI Enterprise support matrix of release 8 requires three things together for an on-premises server: a supported GPU, an NVIDIA-Certified System and supported infrastructure software
- The label states no performance figure, hardware support comes from the system vendors, and NVIDIA's configuration guide sets minimums of two CPU sockets, six physical cores per GPU and system memory of twice the GPU memory
Supplied by Eurokommerz: AI servers, built to order Request a configuration →
What an NVIDIA-Certified System is
An NVIDIA-Certified System is a server, workstation or edge system from a system maker that has passed NVIDIA's certification programme together with named NVIDIA GPUs. NVIDIA's partners, the system makers, run the tests in a standardised setup, NVIDIA engineers evaluate the systems, and NVIDIA publishes the validated models in its "List of NVIDIA-Certified Systems", last updated on 4 October 2026. The label tells a buyer that this model went through NVIDIA's functional and performance tests with the GPUs in its row. It states no performance figure, and hardware support stays with the system vendor.
The sources are NVIDIA's documents as read on 12 November 2026: the programme overview and the configuration guide (both last updated on 30 September 2026), the list page, the product page "NVIDIA-Certified Systems" of 3 September 2026 and the NVIDIA AI Enterprise support matrix of 2 September 2026.
Who tests and who evaluates a certified system
The documents split the work between the system maker and NVIDIA.
Qualification comes first. NVIDIA's product page describes a qualified server as one that "has undergone thermal, mechanical, power, and signal integrity qualification to ensure that the accelerated hardware is fully functional in that server design", and continues: "This testing process is a prerequisite to submitting a system for the NVIDIA-Certified program."
The certification tests are then run by the partner. The configuration guide speaks of "NVIDIA-Certified systems testing by NVIDIA partners" and says that it "requires a standardized setup to simplify performance measurements".
NVIDIA assesses the result. According to the programme overview, certified systems "are evaluated by NVIDIA engineers for performance, functionality, scalability, and security", and the list page introduces its server table with the sentence "The following systems have been validated by NVIDIA as NVIDIA-Certified Systems - Data Center Servers."
Not published are the name of a test tool, pass thresholds, fees, a period of validity and a rule for re-certification.
What the certification tests cover
NVIDIA lists workloads and names no tool. Systems configured as inference reference configurations, workstations, high-density VDI systems and edge devices run workloads such as deep learning training with PyTorch on one and on several GPUs, inference with TensorRT and Triton, data analytics and machine learning with RAPIDS, application development with the CUDA Toolkit and the HPC SDK, and "Inference benchmarking using vLLM". Systems intended for clusters add LLM pretraining benchmarks, multi-node LLM pretraining and a network bandwidth test of RDMA performance.
The scope depends on the category. "NVIDIA-Certified Systems for the data center are tested both as single nodes and in a two-node configuration." Workstations, high-density VDI and edge systems "are evaluated on their standalone performance with the NVIDIA GPUs within a single system". The test environment consists of production versions of Ubuntu 24.04, NVIDIA drivers, the DOCA SDK network software and the NVIDIA Cloud Native Stack with the NVIDIA GPU Operator and Kubernetes; driver and stack versions are not given.
Three further statements matter for a purchase:
- Security. The configuration guide says: "NVIDIA-Certified systems are tested for TPM 2.0 modules." and "NVIDIA-Certified systems are tested for UEFI bootloader compatibility."
- Temperature. Systems "are qualified and tested to run workloads within the OEM manufacturer's temperature and airflow specifications". The limit is the maker's, per model and card; see our guide to GPU server inlet temperature.
- Network. "NVIDIA-Certified Systems for the data center are tested with NVIDIA networking but can be purchased and deployed by customers with their choice of networking."
Servers, workstations and edge systems: the categories
NVIDIA's product page shows three cards: "NVIDIA-Certified Servers", "NVIDIA-Certified Workstations" and "NVIDIA-Certified Edge Systems". The documentation list is divided more finely. Its headings include "Reference Configurations", "Systems Certified for NVIDIA Confidential Computing", "Data Center Servers", "NVIDIA Arm CPU Systems", "Workstations", "Mobile Workstations", "High-Density VDI Servers" and "Edge Systems", among others. Each heading has its own table, so when an offer says "certified", ask under which heading the model stands.
How to read the NVIDIA-Certified Systems list
The list is one long page with a table per category, sorted by partner, and each system name links to the maker's product page. In the table "Data Center Servers" the columns are "Partner", "Data Center Server" and "Supported NVIDIA GPUs". One row is a maker, a model and the GPUs the model is listed with. NVIDIA writes: "Each NVIDIA-Certified System is tested with supported NVIDIA GPUs to validate the performance and reliability of the combined system."
The first block of the page, "Reference Configurations", adds the network. Its tables have the columns "Partner", "System", "Tested NVIDIA GPUs", "Tested Network Devices (East-West)" and "Tested Network Devices (North-South)". The name of a configuration is a code: "Each configuration name follows the pattern CPU-GPU-Network Adaptor-Average GPU Network Bandwidth." NVIDIA's example is 2-4-3-200, which "represents two CPUs, four PCIe GPUs, three network adaptors, and an average network bandwidth of 200 Gb/s".
| CONFIGURATION | NVIDIA'S DESCRIPTION |
|---|---|
| PCIe Optimized 2-2-3-400 | two PCIe GPUs and three network adaptors |
| PCIe Optimized 2-4-3-200 | four PCIe GPUs and three network adaptors |
| PCIe Optimized 2-8-5-200 | eight PCIe GPUs and five network adaptors |
| HGX 2-8-9-400 | NVIDIA HGX H100/H200 systems |
| HGX 2-8-9-800, 2-8-10-800 | NVIDIA HGX B200/B300 systems, average network bandwidth of 800 Gb/s |
| Inference Only 2-8-1, 2-4-1 | RTX PRO 6000 and RTX PRO 4500 Blackwell Server Edition, no East-West network, DPU-based North-South connectivity only |
NVIDIA, "List of NVIDIA-Certified Systems", block "Reference Configurations", last updated 4 October 2026, read 12 November 2026.
For servers with eight PCIe cards, the table "PCIe Optimized 2-8-5-200" held these rows, among others, when it was read for this article.
| SYSTEM | TESTED NVIDIA GPUS |
|---|---|
| Supermicro SYS-422GA-NRT | RTX PRO 6000 Blackwell Server Edition |
| Supermicro SYS-422GL-NR | RTX PRO 6000 Blackwell Server Edition, H200 NVL |
| Supermicro SYS-522GA-NRT | RTX PRO 6000 Blackwell Server Edition, RTX PRO 4500 Blackwell Server Edition, H200 NVL |
| Dell PowerEdge XE7745 | RTX PRO 6000 Blackwell Server Edition, RTX PRO 4500 Blackwell Server Edition, H200 NVL, L40S |
| Dell PowerEdge XE7740 | RTX PRO 6000 Blackwell Server Edition, RTX PRO 4500 Blackwell Server Edition, H200 NVL |
| HPE DL380a Gen12 | RTX PRO 6000 Blackwell Server Edition |
| Gigabyte G494-ZB4-AAP2 | RTX PRO 6000 Blackwell Server Edition |
| ASUS ESC8000A-E13P | L40S |
NVIDIA, "List of NVIDIA-Certified Systems", table "PCIe Optimized 2-8-5-200", rows as read on 12 November 2026; a selection, and the list changes. HPE's full name for its server is HPE ProLiant Compute DL380a Gen12.
The model name alone is not enough. In this table Gigabyte's G494-ZB4-AAP2 stands with the RTX PRO 6000 Blackwell Server Edition, while Gigabyte's own product page for the server also names eight H200 NVL. Whether another table of the list carries that second combination has to be checked on the day of the order. What each maker states for these chassis is compared in our article on eight-GPU servers from Supermicro, Gigabyte and ASUS, and for one card across makers in the list of servers compatible with the RTX PRO 6000 Server Edition.
A row can also carry a condition. The HPE ProLiant Compute DL380 Gen12 stands in the two-GPU table with the RTX PRO 6000 Blackwell Server Edition and a footnote mark, and footnote 1 reads: "GPU power is capped based on system capabilities. Contact the partner for details."
NVIDIA's product page names a second place to search, the Qualified System Catalog on its enterprise marketplace, with the advice to "filter the list for any of the NVIDIA-Certified System categories". That catalogue holds qualified systems as well.
We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project and build AI servers to order; tell us the server model, the cards and whether NVIDIA AI Enterprise support is planned, and we reply with a configuration and quote within one business day.
What the label leaves open
One configuration or every configuration. No NVIDIA document we read says that only the tested configuration is certified, and none says that every configuration of a listed model is. The documents give the row structure, the footnote and the standardised setup. Supermicro is more direct on its own page about NVIDIA-Certified Systems, read on 12 November 2026: "Each server/GPU configuration earns its own certification." In practice, check the row for your model and GPU, and ask the maker about the processor, memory, card count and network adapters of your order.
Performance. The list holds no figure. The programme overview says certified systems "have been proven to deliver predictable performance" and gives no throughput. Published measurements, as in our overview of MLPerf Inference v6.1 results for PCIe GPU servers, answer that for named configurations.
Hardware support. NVIDIA's product page says: "Hardware support for NVIDIA-Certified Systems is available from the system vendors." Who covers which component is the subject of our article on GPU server warranty in the EU.
NVIDIA AI Enterprise and certified hardware
The link between the label and software support is written down in the NVIDIA AI Enterprise documentation. The support matrix of release 8 states: "Supported configurations require all of the following." For on-premises servers it lists a "GPU supported by NVIDIA AI Enterprise (and optionally a supported networking card)", "NVIDIA-Certified Systems" and "Supported infrastructure management software - bare-metal or virtualized (OS, orchestration platform, hypervisor)". The matrix of release 8.2 says the same from the GPU side: "GPUs and accelerated platforms below are supported with servers listed on the NVIDIA-Certified Systems." NVIDIA names one exception: GB200 NVL4, GB200 NVL72 and GB300 NVL72 "do not require NVIDIA-Certified Systems; they require Qualified Systems status".
The certification documentation agrees: "For on-premises deployments (bare metal or virtualized), NVIDIA AI Enterprise is supported on NVIDIA-Certified Data Center and Edge Systems." Workstations do not appear in that sentence. All of these sentences define which configurations NVIDIA supports. None of them says that the software will not start on other hardware.
The GPUs in the 8.2 matrix include the NVIDIA RTX PRO 6000 Blackwell Server Edition, the NVIDIA RTX PRO 4500 Blackwell Server Edition (our pages: buy the RTX PRO 4500 Server Edition and its specifications), the NVIDIA H200 NVL, the L40S and the L4, and among the Ada cards the NVIDIA RTX 5000 Ada Generation (buy page, specifications). The platforms include the NVIDIA DGX B300 and the NVIDIA HGX B300.
Support itself is bought separately: the programme overview says customers "can choose to purchase enterprise-grade support through NVIDIA AI Enterprise Support Services". For virtual machines there is a second list: according to the product page, vGPU software runs on a certified system "as long as this server is also on the list of vGPU-certified servers". Licence terms are covered in our guide to NVIDIA AI Enterprise licensing and, for virtualised hosts, in NVIDIA AI Enterprise on vSphere.
NVIDIA-Certified server requirements in the configuration guide
NVIDIA publishes no entry thresholds for the tests. For buyers it publishes the "NVIDIA-Certified Systems Configuration Guide", which describes itself as "guidelines for configuring NVIDIA-Certified Systems to run various GPU-accelerated computing workloads in production environments". Its table for an inference server contains these lines.
| PARAMETER | NVIDIA'S WORDING |
|---|---|
| GPUs per server | "2x / 4x / 8x GPUs per server for a balance configuration." |
| CPU sockets | "CPU Sockets: minimum 2 sockets." |
| CPU speed | "CPU Speed: minimum 2.0 GHz base clock." |
| CPU cores | "CPU Cores: minimum 6x physical CPU cores per GPU." |
| System memory | "Minimum 2x times of total GPU memory." |
| PCI Express | "RTX Pro 6000 should be connected using a PCIe Gen5 x16 interface or above." |
| Network | "Minimum 200 Gbps for multi-node inference." "Up to 400 Gbps per GPU." |
| Local storage | "one NVMe per CPU socket is recommended"; "Minimum 1TB." |
| Management | "Redfish 1.0 (or greater) compatible" |
| Security | "TPM 2.0 module (secure boot)" |
NVIDIA-Certified Systems Configuration Guide, Table 1 "Inference Server System", last updated 30 September 2026, read 12 November 2026. The guide has the same PCIe line for the H200 NVL and asks for PCIe Gen4 x16 or above for the L40S.
By our arithmetic, a server with eight 96 GB cards holds 768 GB of GPU memory, so the minimums come to 1,536 GB of system memory and 48 physical cores. The training node on our AI servers page follows the same line: eight H200 NVL (8 × 141 GB = 1,128 GB) with 2.3 TB of DDR5 (24 × 96 GB = 2,304 GB) against a minimum of 2,256 GB.
The guide states its own status: "These recommendations serve as a starting point for addressing workload-specific needs." How the lines turn into a request is shown in what to specify before buying an AI server, and how they turn into tender wording in our GPU server tender specification.
Certified, qualified, RTX PRO Server, MGX and HGX
These names appear together in offers. Qualified is the step before certification; NVIDIA adds that "Servers in qualified configurations are supported for production use." NVIDIA's page for the RTX PRO Server, dated 11 June 2026, states that "RTX PRO Servers are NVIDIA-Certified", while the certified list does not use that name and files the systems under the PCIe Optimized and Inference Only reference configurations. MGX is, in NVIDIA's words, "an open modular reference architecture that enables OEMs, ODMs, and ecosystem partners to build accelerated systems faster"; the MGX page does not mention certification, so an MGX server still needs its row in the list.
Makers' systems on the HGX B300 baseboard, such as Supermicro's SYS-822GS-NB3RT and the HPE Compute XD690, stand in the HGX reference tables of the list; our articles on Dell PowerEdge XE servers and HPE ProLiant servers for AI read the makers' documents.
What to ask a supplier before the order
- The exact model name as NVIDIA's list writes it, with every suffix, the table it stands in and the GPU of your order in that row, with its edition.
- Any footnote on the row, such as a GPU power cap, and what the maker says about it.
- The network adapters, if a reference configuration is the target: those tables name the tested devices.
- For NVIDIA AI Enterprise: the supported GPU, the certified system and the operating system or hypervisor in the support matrix, and who provides support.
- Who gives hardware support and warranty for each component, in writing.
- System memory and cores against the configuration guide, and the reason for any deviation.
Whether a server reaches you as a barebone or as a complete tested system changes several of these answers, as our article on barebone and complete GPU servers shows. We size a configuration free of charge and check rack, power and airflow before we quote; send us your models, the number of users and the software stack and we reply with a configuration within one business day.
FAQ
What is an NVIDIA-Certified System?
Where is the NVIDIA-Certified Systems list?
Who certifies a server, NVIDIA or the server maker?
What are the requirements for an NVIDIA-Certified server?
Does NVIDIA AI Enterprise need certified hardware?
Is every configuration of a certified server model certified?
Does NVIDIA support the hardware of a certified system?
We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project, AI servers built to order and NVIDIA AI Enterprise licences across the EU. Describe your task in a few sentences and our team will help. We reply within one business day.
Request a configurationWe reply within one business day