Nvidia DGX, HGX, EGX and AGX - explained and differentiated.

Nvidia is on fire right now, and not just from the perspective of their stock price (343.99 USD 14.54 4.41% today) but from their huge announcements around the future of AI, and several changes to the product offerings on the hardware side.

Perhaps, like me, the plethora of acronyms around their various systems is causing a little confusion. Let us delve into them and help you to understand what the various -X platforms are all about.

Starting small, with AGX. AGX is a platform for AI POCs and Edge Deployments. It is the smallest in physical size of all the -X platforms and was originally just a simple Raspberry-Pi style platform that could be used by developers to test out basic AI inferencing applications. Recently it has morphed into a more complete set of platform offerings, that also includes Nvidia DRIVE AGX for Autonomous Vehicles and Nvidia CLARA AGX for Medical. Start with AGX if you are looking to develop applications and products in these areas.

Many of the AGX platforms are fully designed by Nvidia and use ARM processors alongside Nvidia APUs to enable high levels of performance in a small form factor. They are also available from partners such as Supermicro, Adlink and Advantech. So use these in your robots, kiosks, drones, small retail environments, or in any other small electronic devices that would benefit from AI enablement. They are also perfectly suited to executing small proof of concept experiments in the lab! The link below gives a more comprehensive look into these AGX platforms.

An example Jetson Platform is below and more are pictured here https://www.nvidia.com/en-gb/autonomous-machines/embedded-systems/

Next up the stack is EGX. EGX is a set of validated platforms from OEMs that utilise Nvidia GPUs, and are validated and certified by Nvidia to meet certain performance and functionality requirements. Nvidia work closely with partners and OEMs to ensure these platforms can deliver on promises of high performance across various workloads such as Compute, Design, Virtualisation and AI.

When you think of EGX, think more of server style, rackmount platforms, and some desktop workstations from your favourite vendors like Dell, Supermicro and Gigabyte Enterprise. These platforms ship with Nvidia GPUs integrated and validated, from basic A2 GPU (T4 replacement) right up to A100 in PCIe form factor. Rather than the small form factor units in the AGX line which generally use low power CPUs. These are almost always accompanied by a mainstream datacenter or workstation CPU from one of your favourite vendors, usually Intel or AMD.

Above EGX, in my mind, sits HGX. HGX is the next level up in performance from EGX and is only available with the flagship A100 GPU (now only the 80GB variety). For that reason the HGX platforms are not suited to visualisation - but for compute only.

The Nvidia designed 4 or 8x A100 GPU board sometimes referred to as "Redstone", is once again integrated into 3rd party systems, like the above EGX units. However that is where the similarity with EGX stops. The HGX platform is based around an Nvidia designed (and manufactured?) board with OAM sockets for the top-of-the-line 400w TDP A100. (Outside of HGX you will receive a diluted PCIe A100). With HGX you get 80Gb GPUs with the highest performance, cores and frequencies available on A100. Not only that, you benefit from the Nvidia NVlink technology which connects GPUs directly outside of the PCIe bus allowing for faster data processing and less bottlenecks within the platform.

Link to more info here https://www.nvidia.com/en-gb/data-center/hgx/

Supermicro HGX Platform example:

Finally, the "piece de resistance" or flagship AI platform from Nvidia is the DGX. Now in its 3rd generation, DGX represents the ultimate platform to house your OAM based A100 GPUs. It contains the same "full fat" A100s as most of the HGX platforms on the market, 8 of them - so if the hardware is basically the same as HGX, what is the benefit?

• Nvidia's internal developers and researchers are using these DGX for their own AI work. So all Nvidia learnings on the HW and SW is integrated in the DGX platform.
• Nvidia QA everything from OS to Libraries, AI frameworks and applications on these systems.
• Updates to the open source frameworks, libraries and drivers by Nvidia are immediately available to DGX users.
• The direct support and maintenance with NVIDIA and direct access to NVIDIA DGX experts in case of any issues with the SW environment.

I like to look at DGX as buying a small share in Nvidia and getting full access to the plethora of AI expertise embedded in Nvidia. If you purchase any of the above platforms you get access to a bunch of free resources to help you get off the ground, however if you purchase a DGX, you will be at the cutting edge of AI with significant resource from Nvidia allocated to supporting your AI needs. For me, DGX is really the domain of blue chip companies and research institutes looking to make the next big AI discovery. DGX can be bought on its own, but at this level, consider deployments of multiple DGXs becoming the standard (Pods or Superpods are the given name), in something like an AI centre of excellence at the heart of the business or research strategy.

https://images.nvidia.com/aem-dam/Solutions/Data-Center/nvidia-dgx-a100-datasheet.pdf

If you need more information, please don't hesitate to contact me - I can help you to navigate the Nvidia landscape including the more "normal" GPU cards such as Quadro, or the new A-series. Or I can help you decide on the best hardware for your next AI deployment.

Thanks for reading!