ZXO ROSSO / EXTERIOR CONCEPT
B300.
AI, reimagined.
A bold direction for AI infrastructure.
Meet the ROSSO exterior study and explore the NVIDIA Blackwell Ultra platform that inspires it.
NVIDIA DGX B300
REFERENCE PLATFORM
NEW / ZXO B300 HIGH-DENSITY RACK
ROSSO.
At rack scale.
Four server systems. One black-and-red exterior identity. Explore twelve views of the ZXO B300 high-density rack design study, from the complete cabinet to its material details and wider settings.
Explore the high-density rack
01 / THE PLATFORM
From model training
to AI reasoning.
NVIDIA DGX™ B300 is engineered for enterprise AI environments that bring together large-scale training, inference and reasoning. Built on the NVIDIA Blackwell Ultra architecture, the reference system combines dense GPU compute, substantial memory and high-bandwidth networking.
ZXO ROSSO explores a distinctive exterior identity for this class of computing infrastructure: a sculpted red frame, carbon-black accents and a precise diagonal grille.
Explore NVIDIA DGX B300
02 / THE ROSSO DESIGN LANGUAGE
Precision.
With a pulse.
One selected design, developed across twelve views.
Racing red. Carbon black. A distinct ZXO presence.
A faceted perimeter and diagonal grille establish the ROSSO signature.
Ventilation, service access and rack integration remain part of the engineering brief.
The same exterior language carries from a single node to an illustrative rack deployment.
03 / NVIDIA REFERENCE SPECIFICATIONS
NVIDIA DGX B300
reference platform.
Specifications below describe NVIDIA’s published DGX B300 platform. The ZXO images are exterior concepts.
| GPU architecture | NVIDIA Blackwell Ultra |
|---|---|
| GPU configuration | 8 × NVIDIA Blackwell Ultra SXM GPUs |
| Total GPU memory¹ | 2.1 TB, as listed on NVIDIA’s product page |
| FP8 Tensor Core performance² | 72 PFLOPS sparse / 36 PFLOPS dense |
| FP4 Tensor Core performance² | 144 PFLOPS sparse / 108 PFLOPS dense |
| NVLink aggregate bandwidth | 14.4 TB/s |
| Cluster networking | 8 × ConnectX-8 VPI ports, up to 800 Gb/s each |
| Rack form factor | 10U |
| Platform ecosystem | DGX BasePOD™ and DGX SuperPOD™ reference architectures |
¹ GPU-memory totals differ across NVIDIA publications. Confirm the final system configuration with the supplier.
² Precision and sparsity conditions matter. These are NVIDIA reference-platform figures, not measurements of a ZXO-built system or a delivery commitment.
Sources: NVIDIA DGX B300 product specifications ↗ · NVIDIA datasheet ↗
04 / FROM NODE TO INFRASTRUCTURE
A bigger picture
for your AI project.
Model size, workload, networking, power and cooling shape the deployment conversation. Start with the requirements, then define the system and site plan.
Illustrative deployment scene · Engineering validation pendingSINGLE NODE / HIGH-END REFERENCE
High-End Reference
Configuration.
A single-node planning reference based on NVIDIA DGX B300 hardware, with the upper system-memory option of up to 4 TB. Explore the configuration, then see how four nodes combine in the rack brief.
Planning reference · Engineering validation pending
| Component | Per DGX B300 node |
|---|---|
| GPU platform | 8 × NVIDIA B300 Blackwell Ultra GPUs |
| Nominal GPU-memory capacity¹ | 2,304 GB HBM3e (8 × 288 GB) |
| Host CPUs | 2 × Intel Xeon Platinum 6776P |
| Physical CPU cores | 128 cores (2 × 64) |
| System-memory upper option | Up to 4 TB per node; 2 TB default |
| Boot storage | 2 × 1.92 TB M.2 NVMe / 3.84 TB raw |
| Local data / cache storage | 8 × 3.84 TB E1.S NVMe |
| Raw data / cache capacity² | 30.72 TB |
| Compute-network adapters | 8 × NVIDIA ConnectX-8; up to 800 Gb/s each |
| Storage / management DPUs | 2 × dual-port NVIDIA BlueField-3; 400 Gb/s class |
| NVLink switching | 2 × NVIDIA NVSwitch |
| GPU interconnect domains | One eight-GPU NVLink domain |
| Node-to-node network | InfiniBand / Ethernet compute fabric |
| Reference node envelope | 10U |
¹ GPU capacity follows the hardware guide’s nominal 288 GB per GPU listing. NVIDIA’s DGX B300 product page separately lists 2.1 TB total GPU memory; confirm delivered capacity and the reporting basis with the supplier.
GPU and CPU-memory figures are sums across devices and nodes, and do not establish one unified memory pool. ² Raw NVMe capacities exclude RAID, filesystem and usable-capacity overhead. Data/cache capacity and boot capacity are listed separately.
The ZXO exterior remains a concept. Final components, rack fit and system integration require engineering validation before a delivery specification can be confirmed.
Reference sources: NVIDIA DGX B300 hardware guide ↗ · NVIDIA product specifications ↗ · NVIDIA data-center deployment guide ↗ · Intel Xeon 6776P specifications ↗
View the four-node rack configuration05 / THE COMPLETE VISUAL STUDY
Twelve views.
One direction.
Explore the selected ROSSO design from product views to material details and deployment scenes. Select an image to enlarge it.
All twelve images are AI-generated exterior-concept visualizations. They show design intent, not production hardware, a validated rack build or an actual customer facility.
06 / START WITH THE WORKLOAD
What will you
build with AI?
Share your training or inference workload, model scale, deployment location and infrastructure constraints. Let’s define the right next conversation.