#nvidia dgx

Nvidia HGX vs DGX: What are the differences?

Nvidia is comfortably riding the AI wave. And for at least the next few years, it will likely not be dethroned as the AI hardware market leader. With its extremely popular enterprise solutions powered by the H100 and H200 “Hopper” lineup of GPUs (and now B100 and B200 “Blackwell” GPUs), Nvidia is the go-to manufacturer of high-performance computing (HPC) hardware.

Nvidia DGX is an integrated AI HPC solution targeted toward enterprise customers needing immensely powerful workstation and server solutions for deep learning, generative AI, and data analytics. Nvidia HGX is based on the same underlying GPU technology. However, HGX is a customizable enterprise solution for businesses that want more control and flexibility over their AI HPC systems. But how do these two platforms differ from each other?

Nvidia DGX: The Original Supercomputing Platform

It should surprise no one that Nvidia’s primary focus isn’t on its GeForce lineup of gaming GPUs anymore. Sure, the company enjoys the lion’s share among the best gaming GPUs, but its recent resounding success is driven by enterprise and data center offerings and AI-focused workstation GPUs.

Overview of DGX

The Nvidia DGX platform integrates up to 8 Tensor Core GPUs with Nvidia’s AI software to power accelerated computing and next-gen AI applications. It’s essentially a rack-mount chassis containing 4 or 8 GPUs connected via NVLink, high-end x86 CPUs, and a bunch of Nvidia’s high-speed networking hardware. A single DGX B200 system is capable of 72 petaFLOPS of training and 144 petaFLOPS of inference performance.

Key Features of DGX

AI Software Integration: DGX systems come pre-installed with Nvidia’s AI software stack, making them ready for immediate deployment.

High Performance: With up to 8 Tensor Core GPUs, DGX systems provide top-tier computational power for AI and HPC tasks.

Scalability: Solutions like the DGX SuperPOD integrate multiple DGX systems to form extensive data center configurations.

Current Offerings

The company currently offers both Hopper-based (DGX H100) and Blackwell-based (DGX B200) systems optimized for AI workloads. Customers can go a step further with solutions like the DGX SuperPOD (with DGX GB200 systems) that integrates 36 liquid-cooled Nvidia GB200 Grace Blackwell Superchips, comprised of 36 Nvidia Grace CPUs and 72 Blackwell GPUs. This monstrous setup includes multiple racks connected through Nvidia Quantum InfiniBand, allowing companies to scale thousands of GB200 Superchips.

Legacy and Evolution

Nvidia has been selling DGX systems for quite some time now — from the DGX Server-1 dating back to 2016 to modern DGX B200-based systems. From the Pascal and Volta generations to the Ampere, Hopper, and Blackwell generations, Nvidia’s enterprise HPC business has pioneered numerous innovations and helped in the birth of its customizable platform, Nvidia HGX.

Nvidia HGX: For Businesses That Need More

Build Your Own Supercomputer

For OEMs looking for custom supercomputing solutions, Nvidia HGX offers the same peak performance as its Hopper and Blackwell-based DGX systems but allows OEMs to tweak it as needed. For instance, customers can modify the CPUs, RAM, storage, and networking configuration as they please. Nvidia HGX is actually the baseboard used in the Nvidia DGX system but adheres to Nvidia’s own standard.

Key Features of HGX

Customization: OEMs have the freedom to modify components such as CPUs, RAM, and storage to suit specific requirements.

Flexibility: HGX allows for a modular approach to building AI and HPC solutions, giving enterprises the ability to scale and adapt.

Performance: Nvidia offers HGX in x4 and x8 GPU configurations, with the latest Blackwell-based baseboards only available in the x8 configuration. An HGX B200 system can deliver up to 144 petaFLOPS of performance.

Applications and Use Cases

HGX is designed for enterprises that need high-performance computing solutions but also want the flexibility to customize their systems. It’s ideal for businesses that require scalable AI infrastructure tailored to specific needs, from deep learning and data analytics to large-scale simulations.

Nvidia DGX vs. HGX: Summary

Simplicity vs. Flexibility

While Nvidia DGX represents Nvidia’s line of standardized, unified, and integrated supercomputing solutions, Nvidia HGX unlocks greater customization and flexibility for OEMs to offer more to enterprise customers.

Rapid Deployment vs. Custom Solutions

With Nvidia DGX, the company leans more into cluster solutions that integrate multiple DGX systems into huge and, in the case of the DGX SuperPOD, multi-million-dollar data center solutions. Nvidia HGX, on the other hand, is another way of selling HPC hardware to OEMs at a greater profit margin.

Unified vs. Modular

Nvidia DGX brings rapid deployment and a seamless, hassle-free setup for bigger enterprises. Nvidia HGX provides modular solutions and greater access to the wider industry.

FAQs

What is the primary difference between Nvidia DGX and HGX?

The primary difference lies in customization. DGX offers a standardized, integrated solution ready for deployment, while HGX provides a customizable platform that OEMs can adapt to specific needs.

Which platform is better for rapid deployment?

Nvidia DGX is better suited for rapid deployment as it comes pre-integrated with Nvidia’s AI software stack and requires minimal setup.

Can HGX be used for scalable AI infrastructure?

Yes, Nvidia HGX is designed for scalable AI infrastructure, offering flexibility to customize and expand as per business requirements.

Are DGX and HGX systems compatible with all AI software?

Both DGX and HGX systems are compatible with Nvidia’s AI software stack, which supports a wide range of AI applications and frameworks.

Final Thoughts

Choosing between Nvidia DGX and HGX ultimately depends on your enterprise’s needs. If you require a turnkey solution with rapid deployment, DGX is your go-to. However, if customization and scalability are your top priorities, HGX offers the flexibility to tailor your HPC system to your specific requirements.

NVIDIA, a titan in the world of generative AI accelerators, is not one to rest on its laurels. Despite already dominating the accelerator market, the tech giant is determined to push the envelope further. With the unveiling of its next-generation Blackwell architecture and the B200/B100 accelerators, NVIDIA is set to redefine what’s possible in AI computing yet again.

As we witnessed the return of the in-person GTC for the first time in five years, NVIDIA’s CEO, Jensen Huang, took center stage to introduce an array of new enterprise technologies. However, it was the announcement of the Blackwell architecture that stole the spotlight. This move marks a significant leap forward, building upon the success of NVIDIA’s H100/H200/GH200 series.

Named after the American statistical and mathematical pioneer Dr. David Harold Blackwell, this architecture embodies NVIDIA’s commitment to innovation. Blackwell aims to elevate the performance of NVIDIA’s datacenter and high-performance computing (HPC) accelerators by integrating more features, flexibility, and transistors. This approach is a testament to NVIDIA’s strategy of blending hardware advancements with software optimization to tackle the evolving needs of high-performance accelerators.

With an impressive 208 billion transistors across the complete accelerator, NVIDIA’s first multi-die chip represents a bold step in unified GPU performance. The Blackwell GPUs are designed to function as a single CUDA GPU, thanks to the NV-High Bandwidth Interface (NV-HBI) facilitating an unprecedented 10TB/second of bandwidth. This architectural marvel is complemented by up to 192GB of HBM3E memory, significantly enhancing both the memory capacity and bandwidth.

However, it’s not just about packing more power into the hardware. The Blackwell architecture is engineered to dramatically boost AI training and inference performance while achieving remarkable energy efficiency. This ambition is evident in NVIDIA’s projections of a 4x increase in training performance and a staggering 30x surge in inference performance at the cluster level.

Moreover, NVIDIA is pushing the boundaries of precision with its second-generation transformer engine, capable of handling computations down to FP4 precision. This advancement is crucial for optimizing inference workloads, offering a significant leap in throughput for AI models.

In summary, NVIDIA’s Blackwell architecture and the B200/B100 accelerators represent a formidable advancement in AI accelerator technology. By pushing the limits of architectural efficiency, memory capacity, and computational precision, NVIDIA is not just maintaining its leadership in the AI space but setting new benchmarks for performance and efficiency. As we await the rollout of these groundbreaking products, the tech world watches with anticipation, eager to see how NVIDIA’s latest innovations will shape the future of AI computing.

A frequent topic of inquiry we encounter involves understanding the distinctions between the NVIDIA DGX and NVIDIA HGX platforms. Despite the resemblance in their names, these platforms represent distinct approaches NVIDIA employs to market its 8x GPU systems featuring NVLink technology. The shift in NVIDIA’s business strategy was notably evident during the transition from the NVIDIA P100 “Pascal” to the V100 “Volta” generations. This period marked the significant rise in prominence of the HGX model, a trend that has continued through the A100 “Ampere” and H100 “Hopper” generations.

NVIDIA DGX versus NVIDIA HGX What is the Difference

Focusing primarily on the 8x GPU configurations that utilize NVLink, NVIDIA’s product lineup includes the DGX and HGX lines. While there are other models like the 4x GPU Redstone and Redstone Next, the flagship DGX/HGX (Next) series predominantly features 8x GPU platforms with SXM architecture. To understand these systems better, let’s delve into the process of building an 8x GPU system based on the NVIDIA Tesla P100 with SXM2 configuration.

DeepLearning12 Initial Gear Load Out

Each server manufacturer designs and builds a unique baseboard to accommodate GPUs. NVIDIA provides the GPUs in the SXM form factor, which are then integrated into servers by either the server manufacturers themselves or by a third party like STH.

DeepLearning12 Half Heatsinks Installed 800

This task proved to be quite challenging. We encountered an issue with a prominent server manufacturer based in Texas, where they had applied an excessively thick layer of thermal paste on the heatsinks. This resulted in damage to several trays of GPUs, with many experiencing cracks. This experience led us to create one of our initial videos, aptly titled “The Challenges of SXM2 Installation.” The difficulty primarily arose from the stringent torque specifications required during the GPU installation process.

NVIDIA Tesla P100 V V100 Topology

During this development, NVIDIA established a standard for the 8x SXM GPU platform. This standardization incorporated Broadcom PCIe switches, initially for host connectivity, and subsequently expanded to include Infiniband connectivity.

Microsoft HGX 1 Topology

It also added NVSwitch. NVSwitch was a switch for the NVLink fabric that allowed higher performance communication between GPUs. Originally, NVIDIA had the idea that it could take two of these standardized boards and put them together with this larger switch fabric. The impact, though, was that now the NVIDIA GPU-to-GPU communication would occur on NVIDIA NVSwitch silicon and PCIe would have a standardized topology. HGX was born.

NVIDIA HGX 2 Dual GPU Baseboard Layout

Let’s delve into a comparison of the NVIDIA V100 setup in a server from 2020, renowned for its standout color scheme, particularly in the NVIDIA SXM coolers. When contrasting this with the earlier P100 version, an interesting detail emerges. In the Gigabyte server that housed the P100, one could notice that the SXM2 heatsinks were without branding. This marked a significant shift in NVIDIA’s approach. With the advent of the NVSwitch baseboard equipped with SXM3 sockets, NVIDIA upped its game by integrating not just the sockets but also the GPUs and their cooling systems directly. This move represented a notable advancement in their hardware design strategy.

Consequences

The consequences of this development were significant. Server manufacturers now had the option to acquire an 8-GPU module directly from NVIDIA, eliminating the need to apply excessive thermal paste to the GPUs. This change marked the inception of the NVIDIA HGX topology. It allowed server vendors the flexibility to customize the surrounding hardware as they desired. They could select their preferred specifications for RAM, CPUs, storage, and other components, while adhering to the predetermined GPU configuration determined by the NVIDIA HGX baseboard.

Inspur NF5488M5 Nvidia Smi Topology

This was very successful. In the next generation, the NVSwitch heatsinks got larger, the GPUs lost a great paint job, but we got the NVIDIA A100.

The codename for this baseboard is “Delta”.

Officially, this board was called the NVIDIA HGX.

Inspur NF5488A5 NVIDIA HGX A100 8 GPU Assembly 8x A100 And NVSwitch Heatsinks Side 2

NVIDIA, along with its OEM partners and clients, recognized that increased power could enable the same quantity of GPUs to perform additional tasks. However, this enhancement came with a drawback: higher power consumption led to greater heat generation. This development prompted the introduction of liquid-cooled NVIDIA HGX A100 “Delta” platforms to efficiently manage this heat issue.

Supermicro Liquid Cooling Supermicro

The HGX A100 assembly was initially introduced with its own brand of air cooling systems, distinctively designed by the company.

In the newest “Hopper” series, the cooling systems were upscaled to manage the increased demands of the more powerful GPUs and the enhanced NVSwitch architecture. This upgrade is exemplified in the NVIDIA HGX H100 platform, also known as “Delta Next”.

NVIDIA DGX H100

NVIDIA’s DGX and HGX platforms represent cutting-edge GPU technology, each serving distinct needs in the industry. The DGX series, evolving since the P100 days, integrates HGX baseboards into comprehensive server solutions. Notable examples include the DGX V100 and DGX A100. These systems, crafted by rotating OEMs, offer fixed configurations, ensuring consistent, high-quality performance.

While the DGX H100 sets a high standard, the HGX H100 platform caters to clients seeking customization. It allows OEMs to tailor systems to specific requirements, offering variations in CPU types (including AMD or ARM), Xeon SKU levels, memory, storage, and network interfaces. This flexibility makes HGX ideal for diverse, specialized applications in GPU computing.

Conclusion

NVIDIA’s HGX baseboards streamline the process of integrating 8 GPUs with advanced NVLink and PCIe switched fabric technologies. This innovation allows NVIDIA’s OEM partners to create tailored solutions, giving NVIDIA the flexibility to price HGX boards with higher margins. The HGX platform is primarily focused on providing a robust foundation for custom configurations.

In contrast, NVIDIA’s DGX approach targets the development of high-value AI clusters and their associated ecosystems. The DGX brand, distinct from the DGX Station, represents NVIDIA’s comprehensive systems solution.