NVIDIA DGX
The proven turnkey AI supercomputing platform for building AI factories.
NVIDIA DGX integrates GPUs, CPUs, high-speed networking, storage, and the NVIDIA AI software stack into a turnkey AI supercomputer that is ready to train from day one. The same architecture scales from the desktop (Spark) to 8-GPU servers (B200 · B300), liquid-cooled rack scale (GB200 · GB300 NVL72), data-center-class SuperPOD, and the cloud.
What Is DGX
DGX is a turnkey AI system designed, validated, and supported directly by NVIDIA. It integrates GPUs, Grace/x86 CPUs, NVLink and NVSwitch, InfiniBand/Ethernet networking, storage, and the Base Command and NVIDIA AI Enterprise software into a single product — no component selection or tuning required. Power it on and start large-scale training and inference immediately.
DGX vs HGX — What's the Difference
| Aspect | DGX (Complete System) | HGX (Building Block) |
|---|---|---|
| Form | Turnkey server/rack systems built by NVIDIA | 8-GPU baseboard module (SXM + NVSwitch) |
| Scope | GPU + CPU + networking + storage + software, integrated | GPU and NVLink board only (chassis, CPU, power, cooling by OEM) |
| Supply | Sold and supported directly by NVIDIA and certified partners | Built into OEM servers from Dell, HPE, Supermicro, Cisco, and others |
| Software | Base Command and AI Enterprise pre-integrated and validated | Configured by the OEM or customer |
| Best for | Organizations that need fast deployment and proven reliability | OEMs and large CSPs that want custom server designs |
DGX Lineup (Specs by Line)
DGX Spark
GB10 Grace Blackwell Superchip · 128GB unified memory · 1 PFLOP (FP4) — prototype models up to 200B parameters on the desktop
DGX Station
GB300 Grace Blackwell Ultra desktop · up to 784GB of large unified memory — workstation-class AI development for individuals and teams
DGX B200
8× Blackwell GPUs · 1.4TB HBM3e · 72 PFLOPS training / 144 PFLOPS inference (FP4) · 5th-gen NVLink 1.8TB/s · air-cooled — 3× training and 15× inference vs DGX H100
DGX B300
8× Blackwell Ultra GPUs · expanded HBM3e (~2.3TB class) · air-cooled 8-GPU — the higher-tier generation optimized for large-scale inference and agentic/reasoning AI
DGX GB200 NVL72
A liquid-cooled rack connecting 72 Blackwell GPUs + 36 Grace CPUs over NVLink as one giant GPU · ~1.4 EFLOPS training — up to 30× H100 for real-time trillion-parameter LLM inference
DGX GB300 NVL72
72 Blackwell Ultra + 36 Grace · rack-scale high-capacity memory · liquid-cooled — built for inference factories with dramatically higher inference throughput
DGX SuperPOD
Data-center-scale turnkey infrastructure linking many DGX systems (B200 · GB200 · GB300) over Quantum-X800 InfiniBand · unified operations with Base Command
DGX Cloud
DGX infrastructure as a service on AWS, Azure, Google Cloud, and OCI — large-scale training and inference immediately, with no upfront investment
Next-Generation Roadmap — Vera Rubin · Feynman
DGX (Vera Rubin · Rubin NVL144)
2026 next generation — Vera CPU + Rubin GPU · HBM4 · NVLink 6 · 144-GPU domain per rack · ~3.6 EFLOPS FP4 inference (roadmap). Scales further with Rubin Ultra NVL576 (2027)
Feynman
Planned for 2028 — the architecture generation after Rubin. Pairs with the Vera CPU and next-generation HBM and NVLink to take AI factory performance up another level (per roadmap announcement; detailed specs to follow)
Key Spec Comparison
| Model | Architecture | GPU Configuration | Memory | AI Performance (FP4 Inference) | Cooling · Form Factor | Positioning |
|---|---|---|---|---|---|---|
| DGX B200 | Blackwell | 8× B200 | 1.4TB HBM3e | 144 PFLOPS | Air-cooled · server | Standard training & inference |
| DGX B300 | Blackwell Ultra | 8× B300 | ~2.3TB HBM3e | ≈216 PFLOPS * | Air-cooled · server | Large-scale inference & agentic AI |
| DGX GB200 NVL72 | Blackwell | 72× B200 + 36 Grace | Rack scale | 1.44 EFLOPS | Liquid-cooled · rack | Trillion-parameter LLMs |
| DGX GB300 NVL72 | Blackwell Ultra | 72× B300 + 36 Grace | Rack scale (high capacity) | ≈1.5 EFLOPS * | Liquid-cooled · rack | Inference factories |
| DGX Rubin NVL144 | Vera Rubin (2026) | 144× Rubin + Vera | HBM4 | ≈3.6 EFLOPS * | Liquid-cooled · rack | Next-generation roadmap |
* AI performance figures are NVIDIA-published values based on FP4 inference (PFLOPS/EFLOPS). Values marked * are approximate or roadmap-based and may differ from final specs. (Training examples — DGX B200: FP8 72 PFLOPS · GB200 NVL72: FP8 720 PFLOPS)