AIHealth
On-premise clinical platform with local LLMs, RAG on FHIR/DICOM data, diagnostic support, remote follow-up. Architecture designed for the MDR pathway.
Discover AIHealth →
Artificial Intelligence
EU AI Act consulting: system classification, policy definition, AI governance, training.
Discover →DGX-1 (2016): turnkey AI server
The DGX programme was NVIDIA’s response to a specific demand from enterprises and research labs: a turnkey deep-learning training system, pre-integrated at the hardware, software and support levels. The first system, DGX-1, was announced in April 2016 (GTC, Jensen Huang):
- 8 Tesla P100 GPUs (Pascal, 16 GB HBM2) interconnected in a hybrid cube-mesh via NVLink 1.0
- 2 Intel Xeon CPUs, 512 GB RAM, 8 TB NVMe SSD
- Pre-installed stack: CUDA driver, cuDNN, DL frameworks (Caffe, TensorFlow, Torch, Theano), NVIDIA GPU Cloud
The 2017 DGX-1 V100 refresh upgraded the GPUs to V100 (Volta, first-generation Tensor Cores).
DGX-2 and NVSwitch (2018)
DGX-2 (March 2018) introduced NVSwitch, a crossbar chip enabling 16 V100 GPUs to communicate all-to-all at 300 GB/s bidirectional per GPU, removing the mesh-topology bottleneck of DGX-1. Aggregate memory reached 512 GB HBM2. DGX-2 was the first system to expose 512 GB of unified GPU memory to a single training job.
DGX A100 and DGX H100
- DGX A100 (2020) — 8× A100 (40 GB, later 80 GB), 6× NVSwitch, 8× ConnectX-6 VPI (HDR InfiniBand 200 Gb/s), 2× AMD EPYC Rome. Introduced MIG partitioning of each A100 into 7 isolated instances.
- DGX H100 (2022) — 8× H100 (80 GB HBM3), NVSwitch 3, 4× ConnectX-7 (NDR 400 Gb/s), 2× Intel Sapphire Rapids. Rated at ~32 petaFLOPS in FP8 per system.
On the software side, each DGX generation is supported by DGX OS (Ubuntu Server with pre-qualified NVIDIA patches and drivers) and by NGC (NVIDIA’s container registry for PyTorch, TensorFlow, Triton, RAPIDS).
DGX GH200 (2023)
At GTC on 28 May 2023 (Computex), NVIDIA announced DGX GH200, a rack-cluster scale configuration:
- 256 GH200 Grace Hopper Superchip (ARM Grace CPU + H100 GPU on a single module, linked via NVLink-C2C at 900 GB/s)
- NVLink Switch System with a two-level fat-tree topology
- 144 TB of unified memory addressable as a single space (HBM3 + LPDDR5X)
- Interconnect bandwidth ~128 TB/s
The stated target is training of memory-bound models (graphs, recommender systems, LLMs with very large embedding tables) without manual memory sharding.
DGX GB200 NVL72 (GTC 2024)
At GTC on 18 March 2024, alongside Blackwell, NVIDIA unveiled GB200 NVL72 as a standard rack unit:
- 36 Grace CPUs + 72 B200 per rack (18 compute trays × 2 GB200 Superchips)
- NVLink 5 and 4th-generation NVLink Switch, 1.8 TB/s NVLink bandwidth per GPU
- Liquid-cooled throughout the rack
- ~1.4 exaFLOPS in FP4, ~720 petaFLOPS in FP8
NVL72 is designed for training and serving of trillion-parameter models, with an NVLink domain extended to 72 GPUs (versus 8 in DGX H100).
GB300 NVL72 and DGX Station (GTC 2025)
At GTC on 18 March 2025, NVIDIA presented the Blackwell Ultra generation and the associated DGX platforms:
- GB300 Superchip: refresh of GB200 with B300 GPUs (Blackwell Ultra), HBM3e up to 288 GB per GPU and higher stated FP4 throughput than B200
- GB300 NVL72: same rack form factor as GB200 NVL72, with 72 B300 GPUs + 36 Grace CPUs and NVLink 5
- DGX Station (2025): desktop workstation based on a single GB300 Superchip, up to 784 GB of coherent memory (HBM3e + LPDDR5X), liquid cooling in a tower chassis
- DGX Spark (previously announced as Project DIGITS): mini-workstation with GB10 (low-end Blackwell + Grace ARM), 128 GB of unified memory, aimed at individual developers and small teams. The architectural pattern — ARM CPU + GPU + unified DRAM on a single module — was first introduced to the consumer/workstation segment by Apple Silicon (M1, November 2020); NVIDIA reuses it here in an AI-focused configuration
In parallel NVIDIA published the next roadmap — Vera Rubin as the post-Blackwell architecture codename, scheduled for subsequent cycles.
DGX SuperPOD
DGX SuperPOD is the reference architecture for multi-system clusters: DGX units connected by InfiniBand (first HDR, then NDR/XDR) in fat-tree or dragonfly+ topology, high-performance storage (NVIDIA GPUDirect Storage, WekaIO, DDN, VAST Data), orchestration via Slurm or Base Command Manager. Reference sizes range from 32 to 1024+ DGX. Known deployments: Meta AI RSC, NVIDIA Eos, EuroHPC clusters.
DGX Cloud
DGX Cloud (announced at GTC in March 2023) is the managed offering: DGX capacity hosted on hyperscalers (Oracle OCI, Microsoft Azure, Google Cloud, AWS) and exposed as an NVIDIA service with Base Command and AI Enterprise included. Customers rent monthly capacity without owning the hardware. Relevant for spiky training workloads or proof-of-concept at scale.
Positioning and costs
A DGX H100 has had historic list pricing around ~€400-500k. An NVL72 is a multi-million-euro unit, aimed at hyperscalers, national labs and organisations building frontier models. For SMEs and research centres the typical path is DGX Cloud or clusters of individual H100 PCIe / HGX H100 (8-GPU OEM baseboards) mounted in Supermicro/Dell/Lenovo servers — same silicon, lower price, but the NVLink domain is limited to 8 GPUs.
The noze context
In R&D, noze does not operate DGX in-house but designs AIHealth to run on NVIDIA-compatible stacks: H100 / H200 on OEM servers in colocation, or professional RTX GPUs for inference and fine-tuning of mid-sized models (≤ 70 B quantised parameters). The choice between DGX Cloud, on-premise HGX and workstation-class cards depends on workload profile (training vs inference) and on healthcare-data residency constraints.
References: DGX-1 announced at GTC keynote 5 April 2016 (Jensen Huang). DGX-2 announced at GTC 27 March 2018, first platform with NVSwitch. DGX A100 announced 14 May 2020 (GTC Digital). DGX H100 announced 22 March 2022 (GTC Spring). DGX GH200 announced at Computex, 28 May 2023. DGX Cloud announced at GTC, 21 March 2023. GB200 NVL72 presented at GTC, 18 March 2024. GB300 NVL72, DGX Station and DGX Spark announced at GTC on 18 March 2025 (Jensen Huang keynote). Sources: NVIDIA press releases, nvidia.com/dgx product pages, NVLink and NVSwitch whitepapers.