SERVERS

Discover the architecture that powers large language models like ChatGPT by integrating our servers, equipped with NVIDIA’s most powerful GPUs, into your business processes.

OpenZeka – NVIDIA DGX AI Compute Systems Partner

GPU Servers are designed for artificial intelligence, high-performance computing (HPC), data science, and advanced analytics. While PCIe-based solutions provide cost efficiency and broad compatibility, SXM-based DGX and HGX systems deliver superior performance for large-scale AI training with NVLink and NVSwitch technologies, offering low latency and high bandwidth. The SXM architecture forms the backbone of modern large language models such as ChatGPT, Claude, Gemini, and LLaMA.

DGX Solutions HGX Solutions PCIe Solutions
Description GPU servers optimized end-to-end by NVIDIA, combining GPU, CPU, memory, storage, and networking. Flexible GPU servers offered by OEM vendors, built on NVIDIA’s SXM GPU and NVSwitch infrastructure, with additional CPU, memory, storage, and networking components. Cost-effective and compatible servers using standard PCIe GPU cards.
Manufacturer NVIDIA OEM vendors (HGX platform provided by NVIDIA) OEM vendors (Dell, HP, Supermicro, Gigabyte, etc.
Performance Direct GPU-to-GPU communication ensures low latency and high bandwidth for maximum performance. Direct GPU-to-GPU communication ensures low latency and high bandwidth for maximum performance. Since GPUs communicate over the PCIe bus, performance is lower compared to DGX and HGX solutions.
Flexibility Limited flexibility, as NVIDIA delivers fully configured systems with fixed hardware setups. OEM vendors can offer various combinations of CPU, RAM, storage, and networking. The number of GPUs can vary depending on the model (e.g., 4 or 8 SXM GPUs). High flexibility—GPUs can be added or removed easily thanks to PCIe design, and systems can be scaled according to specific needs.
Scalability NVSwitch + InfiniBand enable scaling up to thousands of GPUs. NVSwitch + InfiniBand enable scaling up to thousands of GPUs. Typically limited to 4–8 GPUs within a single server.
Use Cases Large language models, HPC, advanced research. Large language models, HPC, advanced research, and customized enterprise solutions. Mid-scale training, inference, data analytics, and visualization.

DGX SOLUTIONS

DGX Solutions are GPU servers fully integrated by NVIDIA, optimized for AI and HPC. With easy deployment and maximum performance, they are ready to be applied directly to large-scale workloads.

NVIDIA DGX B300

AI system powered by NVIDIA Blackwell Ultra for training and inference of large generative AI and other transformer-based workloads.

Learn More>>

NVIDIA DGX B200

Unified AI system built with NVIDIA Blackwell for every stage of the AI pipeline, from training to fine-tuning to inference.

Learn More >>

Learn More >>

NVIDIA DGX H200

AI supercomputer optimized for large generative AI and other transformer-based workloads.

Learn More >>

Learn More >>

NVIDIA DGX BasePOD

Proven reference architecture for AI infrastructure delivered by leading storage providers.

Learn More >>

NVIDIA DGX SuperPOD

Leadership-class AI infrastructure for on-premises and hybrid deployments, configurable with any NVIDIA DGX system.

Learn More >>

Learn More >>

NVIDIA Mission Control

Full-stack intelligence that simplifies AI operations with agility, resilience, and hyperscale efficiency for enterprises.

 

HGX SOLUTIONS

HGX Solutions are built by integrating NVIDIA’s HGX platform into server infrastructures by OEM vendors. This creates flexible and high-performance GPU servers that can be adapted to various workloads such as AI training, HPC, and data analytics.

HGX B300 HGX B200 HGX H200
4-GPU 8-GPU
Form Factor 8x NVIDIA Blackwell Ultra SXM 8x NVIDIA Blackwell SXM 4x NVIDIA H200 SXM 8x NVIDIA H200 SXM
FP4 Tensor Core 144 PFLOPS | 108 PFLOPS 144 PFLOPS | 72 PFLOPS 16 PFLOPS 32 PFLOPS
TF32 Tensor Core 18 PFLOPS 18 PFLOPS 4 PFLOPS 8 PFLOPS
Total Memory Up to 2.1 TB 1.4 TB 564 GB HBM3e 1.1 TB HBM3e
Networking Bandwidth 1.6 TB/s 0.8 TB/s 0.4 TB/s 0.8 TB/s

PCIe SOLUTIONS

PCIe Solutions are cost-effective and flexible server solutions that use standard PCIe GPU cards. Thanks to their high compatibility, they can be easily integrated with different systems and deliver balanced performance for mid-scale AI training, inference, data analytics, and visualization workloads.

Frequently Asked Questions (FAQ)

  • SXM GPUs enable direct, low-latency, high-bandwidth communication between GPUs using NVLink and NVSwitch technologies. This architecture allows up to 8 SXM GPUs to operate in a full-mesh connection within a single server and scale to thousands of GPUs across hundreds of servers via InfiniBand.
  • PCIe GPUs, on the other hand, follow the standard card format, where GPU-to-GPU communication occurs indirectly over the PCIe bus. As a result, latency is higher and bandwidth is limited. Typically restricted to 4–8 GPUs per server, PCIe-based systems are significantly less efficient than SXM solutions when scaling across multiple servers.

No. SXM and PCIe are different physical form factors and are not compatible with each other. Even if they share the same GPU architecture (e.g., H100), the SXM version connects directly to the motherboard via a socket, while the PCIe version is installed as a card into a PCIe slot. Therefore, one cannot be used in place of the other.

  • DGX is a turnkey GPU server solution manufactured by NVIDIA, delivered with all components such as CPU, memory, storage, and networking. Since it is fully optimized by NVIDIA, it is ready to use right out of the box.
  • HGX, on the other hand, is a platform where NVIDIA provides only the GPU and NVSwitch infrastructure. OEM vendors (such as Dell, Supermicro, Gigabyte, etc.) complete this with CPU, memory, storage, and networking components to deliver ready-to-deploy GPU servers with more flexible configurations.

In terms of GPU performance, there is no difference since both use the same SXM GPUs and NVSwitch infrastructure. However, overall system performance may vary depending on the CPU, memory, storage, and networking components used. DGX servers deliver predictable performance with NVIDIA’s fixed and optimized hardware, while HGX-based servers can offer lower or higher performance depending on the OEM vendor’s configuration choices.

PCIe-based servers can be used for training small to medium-scale models. A single server typically supports 4–8 GPUs, which may be sufficient for mid-scale AI projects. However, training very large language models such as ChatGPT, Claude, Gemini, or LLaMA—each with hundreds of billions of parameters—requires SXM-based DGX and HGX systems. These systems enable thousands of GPUs to operate efficiently in low-latency, full-mesh configurations. PCIe servers are not efficient at this scale.

Yes. NVIDIA manufactures a specialized server board called the HGX board, which includes SXM GPUs along with the NVSwitch infrastructure that enables GPU-to-GPU communication. OEM vendors (such as Dell, Supermicro, Gigabyte, etc.) integrate this board with additional components like CPU, memory, storage, networking, and cooling to deliver a fully equipped, ready-to-use server. Therefore, when you order an HGX-based server, the GPUs are already included and pre-installed.

  • PCIe-based servers: Yes. Thanks to the standard card format, GPUs can be easily removed and upgraded with newer models. This flexibility allows the system to quickly adapt to new workloads while protecting your investment.
  • SXM-based servers: No. GPUs come integrated through SXM sockets on the motherboard and cannot be replaced.