NVIDIA DGX Systems
Purpose-built AI training systems — “the gold standard for AI infrastructure.”
DGX Systems Lineup
| System | GPUs | Key positioning |
|---|---|---|
| DGX H100/H200 | 8× H100/H200 SXM | Optimized for AI and all transformer-based workloads |
| DGX B200 | 8× B200 SXM | Built with NVIDIA Blackwell; from training to inference and all other transformer-based workloads |
| DGX B300 | 8× B300 SXM | Latest Blackwell Ultra; highest AI throughput per node |
| DGX GB200 | Multiple GB200 | Leadership-class training; foundational model training |
| DGX GB300 NVL72 | 72× B300 + 36× Grace | NVIDIA’s most advanced Superchip AI server rack |
DGX H100 — The Gold Standard
“The world’s first AI system with the power of a data center.”
| Component | Specification |
|---|---|
| GPUs | 8× NVIDIA H100 SXM5 80GB Tensor Core GPUs |
| GPU Interconnect | 4× NVSwitch (all-to-all 900 GB/s per GPU) |
| System Memory | 2 TB |
| Storage | 30 TB NVMe SSDs |
| Network | 8× ConnectX-7 400 Gbps InfiniBand NICs |
| CPUs | 2× Intel Xeon Platinum 8480C |
| AI Performance | 32 quadrillion operations per second |
DGX H100 Physical Design
- Form factor: 10U rack unit
- Components: Front panel (power, I/O), fan modules, NVLink switches, 8× H100 GPUs, NVMe drives, power supplies, 4× ConnectX-7 network modules, motherboard tray
GB300 NVL72 — Rack-Scale AI
“Built for the age of AI reasoning.”
The GB300 NVL72 is not a single server — it’s an entire rack as one computing unit:
| Component | Specification |
|---|---|
| GPUs | 72× NVIDIA Blackwell Ultra (B300) |
| CPUs | 36× NVIDIA Grace CPUs |
| NVLink generation | 5th generation |
| NVLink Bandwidth | 130 TB/s total |
| GPU Memory | 20 TB total @ 576 TB/s bandwidth |
| CPU Memory | 14 TB LPDDR5X |
Key concept: All 72 GPUs are connected via 5th-gen NVLink — they form a single, unified compute fabric at rack scale. There is no InfiniBand “hop” between nodes within the rack; NVLink connects everything.
RTX PRO Server
“The Most Powerful Blackwell Data Center Platform for AI and Visual Computing”
NVIDIA RTX PRO 6000 Blackwell Server Edition GPU:
| Spec | Value |
|---|---|
| CUDA parallel processing cores | 24,064 |
| Tensor Cores | 752 (5th generation) |
| RT Cores | 188 (4th generation) |
| Architecture | Blackwell |
| Target | AI workloads + professional visualization in data center |
The RTX PRO Server is the platform for organizations that need both high-performance AI inference and photorealistic rendering, simulation, or digital twin workloads on the same server infrastructure.
DGX vs NVIDIA-Certified Servers
| DGX Systems | NVIDIA-Certified Servers | |
|---|---|---|
| Built by | NVIDIA | Third-party OEMs (Dell, HPE, Lenovo, etc.) |
| Configuration | Fixed, NVIDIA-engineered | Variable, NVIDIA-validated |
| Support | NVIDIA Support (optional) | OEM support |
| Use case | Maximum performance; NVIDIA reference design | Flexible deployments, existing vendor relationships |
| GPU options | NVIDIA SXM (highest perf) | PCIe or SXM depending on OEM |