NVIDIA launched its DGX range of AI servers back in 2016. Since then there’s been multiple generations of DGX systems plus HGX, MGX and EGX variants. So what’s the difference between these and why does DGX still offer the best AI experience?
EGX servers are based on NVIDIA RTX PRO PCIe GPU accelerators and are available in a wide range of custom configurations offering flexibility when it comes to CPU, memory, storage and networking choices. In contrast MGX servers are largely based around the NVIDIA Superchip combining a CPU and GPU in a single die. This offers less flexibility but great scalability for datacentre environments. HGX servers feature SXM format GPU accelerators in a pre-defined eight GPU tray, offering the same GPU specifications as a DGX but with the option of CPU, memory, storage and network choice once again. Finally the DGX is a fixed configuration AI appliance featuring eight SXM GPUs, twin CPUs, fixed memory and storage capacity and NVIDIA networking.
EGX
MGX
HGX
DGX
So why is the less flexible DGX best? This comes down to the DGX being the only purpose-built AI appliance among all the systems available. And as an appliance the DGX is only AI solution that combines the best NVIDIA hardware and software, plus specialist technical support from NVIDIA DGXperts.
In terms of hardware, the GPUs, CPUs and system memory are selected and matched for optimum performance and linked via NVLink and NVSwitch to get the maximum throughput and it is the only one of the four system types to come preinstalled with NVIDIA AI Enterprise software stack and dedicated management software. This means all of NVIDIA’s AI expertise is available and pre-optimised for a DGX, making the DGX a true turnkey appliance, without any additional software needs. There are also media retention packages directly with NVIDIA.
DGX SOFTWARE STACK
AI ENTERPRISE SOFTWARE STACK
As these critical software platforms are included with the DGX appliances, so are updates, revisions and new features are being added all the time. So when you buy a DGX you’re bound to see performance uplifts over time as new frameworks and applications are released. Only the DGX provide this combination of hardware, software and services all working in unison.
What's more, you can easily cluster multiple DGXs together into what's known as DGX BasePODs or SuperPODs, with as many as 140 nodes acting together as the building blocks for AI factories. Using DGXs to build large AI clusters like this is simpler than any other system as NVIDIA provides reference architectures and software such as Unified Fabric Manager.
And for a similar development experience, there's the DGX SPARK and DGX Station too - both desktop devices, ideal for getting AI projects both small and large, off the ground. The DGX Spark shares the same software platform with its larger datacentre-grade DGX's, so the NVIDIA AI ecosystem now spans software and hardware across all three stages of AI project: development, training and inferencing. This common software base also accelerates projects and eases support.
