Ateamdata / Hardware reference

NVIDIA GPU matrixFrom architecture to system.

Explore data-center GPU families across the formats you can specify. Find a model, compare the layers, and open an entry for context and its official source.

H100 vs H200 guide

8 families & refreshes5 product formats27 representative entriesReviewed Oct 29, 2026

Representative NVIDIA data-center GPU families by product format. Expand an entry for details and manufacturer sources.
Architecture↓ Family / Format →	01 / AcceleratorPCIe cards	02 / Module + platformSXM / HGX	03 / SuperchipCPU + GPU	04 / Complete systemDGX servers	05 / Rack-scale systemNVL72 racks
RubinVera + Rubin platforms	No example included in this guide	HGX Rubin NVL88 × Rubin SXM GPUs An HGX platform with eight Rubin GPUs. The Vera CPU configuration is named HGX Vera Rubin NVL8; x86-based CPU baseboards are another option. NVIDIA reference for HGX Rubin NVL8 (opens in a new tab)	No example included in this guide	No example included in this guide	Vera Rubin NVL7272 GPUs · liquid-cooled rack The DGX Vera Rubin rack contains 18 compute trays, each with two Vera CPUs and four Rubin GPUs, plus NVLink switch trays. NVIDIA reference for Vera Rubin NVL72 (opens in a new tab)
Blackwell UltraBlackwell family refresh	No example included in this guide	HGX B3008 × Blackwell Ultra SXM GPUs A multi-GPU HGX baseboard for systems built around Blackwell Ultra GPUs and NVLink 5 switches. NVIDIA reference for HGX B300 (opens in a new tab)	GB300 SuperchipGrace + Blackwell Ultra The rack-system Grace Blackwell Ultra Superchip couples one Grace CPU with two Blackwell Ultra GPUs through NVLink-C2C. The desktop variant is outside this guide. NVIDIA reference for GB300 Superchip (opens in a new tab)	DGX B3008 × B300 GPUs An NVIDIA DGX server built around eight B300 Blackwell Ultra GPUs. NVIDIA reference for DGX B300 (opens in a new tab)	GB300 NVL7272 GPUs · 36 Grace CPUs A liquid-cooled rack-scale platform integrating Blackwell Ultra GPUs, Grace CPUs and a 72-GPU NVLink domain. NVIDIA reference for GB300 NVL72 (opens in a new tab)
BlackwellCompute + RTX branches	RTX PRO 6000Blackwell Server Edition The Server Edition is a PCIe GPU for enterprise AI and visual computing. Do not substitute the Workstation Edition or Max-Q variant when specifying a server. NVIDIA reference for RTX PRO 6000 (opens in a new tab)	HGX B2008 × Blackwell SXM GPUs An HGX baseboard combining eight Blackwell GPUs with NVLink 5 switches for manufacturer server platforms. NVIDIA reference for HGX B200 (opens in a new tab)	GB200 SuperchipGrace + Blackwell One Grace CPU and two Blackwell GPUs are connected through NVLink-C2C. This assembly is a building block within a larger system. NVIDIA reference for GB200 Superchip (opens in a new tab)	DGX B2008 × B200 GPUs An NVIDIA DGX server built around eight B200 Tensor Core GPUs. NVIDIA reference for DGX B200 (opens in a new tab)	GB200 NVL7272 GPUs · 36 Grace CPUs A liquid-cooled rack-scale system combining Grace CPUs and Blackwell GPUs in a 72-GPU NVLink domain. NVIDIA reference for GB200 NVL72 (opens in a new tab)
HopperH100 / H200 familyExplore our servers ↗	H100 PCIePCIe accelerator card The PCIe add-in-card variant of H100. It is distinct from H100 SXM and H100 NVL; use the exact variant in a bill of materials. NVIDIA reference for H100 PCIe (opens in a new tab) H100 NVLPCIe · NVLink-connected GPUs H100 NVL uses PCIe cards. NVL in this product name does not mean a complete NVL72 rack. NVIDIA reference for H100 NVL (opens in a new tab) H200 NVLPCIe · NVLink bridge options The PCIe variant of H200 supports two- or four-way NVLink bridges. It is distinct from H200 SXM and from a rack-scale NVL72 system. NVIDIA reference for H200 NVL (opens in a new tab)	H100 SXM / HGX H100SXM GPUs · 4 / 8-GPU HGX H100 SXM GPUs are used in HGX H100 partner platforms. Our featured H100 server configurations use eight SXM GPUs. NVIDIA reference for H100 SXM / HGX H100 (opens in a new tab) H200 SXM / HGX H200SXM GPUs · 4 / 8-GPU HGX H200 SXM GPUs are used in HGX H200 partner platforms. Our featured H200 server configurations use eight SXM GPUs. NVIDIA reference for H200 SXM / HGX H200 (opens in a new tab)	GH200 SuperchipGrace + Hopper Grace Hopper combines a Grace CPU and a Hopper GPU through NVLink-C2C. HBM3 and HBM3e configurations exist; specify the intended variant. NVIDIA reference for GH200 Superchip (opens in a new tab)	DGX H1008 × H100 GPUs An NVIDIA DGX server with eight H100 GPUs. A manufacturer server built on HGX H100 is a different system product. NVIDIA reference for DGX H100 (opens in a new tab) DGX H2008 × H200 GPUs An NVIDIA DGX server with eight H200 GPUs. A manufacturer server built on HGX H200 is a different system product. NVIDIA reference for DGX H200 (opens in a new tab)	No example included in this guide
Ada LovelaceGraphics + inference branch	L4Low-profile · single-slot PCIe An Ada Lovelace GPU for AI, video and visual computing in a low-profile PCIe form factor. NVIDIA reference for L4 (opens in a new tab) L40SDual-slot PCIe An Ada Lovelace GPU for AI and graphics workloads. L40S uses PCIe and does not support NVLink. NVIDIA reference for L40S (opens in a new tab)	No example included in this guide	No example included in this guide	No example included in this guide	No example included in this guide
AmpereA100 family	A100 PCIePCIe accelerator card The PCIe form of the Ampere A100 GPU. Confirm the memory variant and server compatibility when specifying a system. NVIDIA reference for A100 PCIe (opens in a new tab)	A100 SXM / HGX A100SXM GPUs · 4 / 8-GPU HGX The SXM form of A100 is used on HGX A100 baseboards. It is a different physical format from the PCIe card. NVIDIA reference for A100 SXM / HGX A100 (opens in a new tab)	No example included in this guide	DGX A1008 × A100 GPUs An NVIDIA DGX system containing eight Ampere A100 GPUs. NVIDIA reference for DGX A100 (opens in a new tab)	No example included in this guide
TuringT4 family	T4Small-form-factor PCIe A Turing GPU for inference, video and graphics workloads, packaged as a 70-watt PCIe accelerator. NVIDIA reference for T4 (opens in a new tab)	No example included in this guide	No example included in this guide	No example included in this guide	No example included in this guide
VoltaV100 family	V100 / V100S PCIePCIe accelerator cards Volta-generation PCIe accelerators. V100 and V100S are distinct variants; check the exact memory and card specification. NVIDIA reference for V100 / V100S PCIe (opens in a new tab)	V100 SXM2SXM2 GPU module The SXM2 module variant of V100. It requires a compatible platform and is not interchangeable with a PCIe add-in card. NVIDIA reference for V100 SXM2 (opens in a new tab)	No example included in this guide	No example included in this guide	No example included in this guide

— No example included in this curated guide; it does not mean a product cannot exist. This is a reference, not a stock or availability listing.

Scroll sideways to compare formats ↔

Read the layers

One architecture.
Different things to buy.

The columns span components, platforms and complete systems. They are not interchangeable products or a performance ranking.

PCIe cards

Add-in GPUs for compatible servers. Check the exact card, power, cooling and server qualification.

SXM / HGX

SXM is a GPU module format. HGX combines GPUs and interconnects into a platform for manufacturer systems.

CPU + GPU

A tightly connected CPU–GPU assembly, such as Grace Hopper or Grace Blackwell, used inside a larger system.

DGX servers

Selected NVIDIA DGX server systems. DGX also includes other product types; rack systems are shown in the next column.

NVL72 racks

Integrated 72-GPU NVLink domains with compute, switching and liquid cooling. These require facility-level planning.

A few names worth separating.

Hopper & Ada

These are parallel product families with different focus areas. Ada is not presented here as the successor to Hopper. Blackwell Ultra is shown separately as a refresh within the Blackwell family.

NVL does not always mean a rack

H100 NVL and H200 NVL belong in the PCIe column. GB200 NVL72, GB300 NVL72 and Vera Rubin NVL72 are rack-scale systems. HGX Rubin NVL8 belongs in the HGX column.

Where does MGX fit?

MGX is a modular reference architecture used by system manufacturers across GPU families and system sizes. It is not a GPU generation or a single card format.

NVIDIA MGX reference ↗

Scope: selected server and data-center products, from Volta to Rubin. GeForce, desktop workstations, embedded platforms and full cluster designs are outside this table. Product names and configurations follow the linked NVIDIA references as of the review date.

From reference to requirements

Found your starting point?

Explore our H100 / H200 server platforms or discuss a sourcing and integration brief.

Explore server platforms