NVIDIA HGX B300
Eight Blackwell Ultra GPUs with up to 288 GB HBM3e per GPU. 50% more memory than B200 and over 100 PFLOPS dense FP4 per node. Built for AI reasoning at scale. Designed, deployed, and managed by CUDO.
NVIDIA HGX B300
Infrastructure and technology partners
Perfect for a range of workloads
Over 250 tokens per second per user on DeepSeek R1-671B from a single 8-GPU node. Up to 2.3 TB of HBM3e per node keeps entire models in memory for low-latency inference on dedicated B300 clusters.
Up to 1.5X faster training than B200 with FP4 precision. Up to 2.3 TB HBM3e per node for trillion-parameter models and large MoE architectures with fewer memory bottlenecks.
Full model weights held in memory for agentic workflows and long-context reasoning. No offloading, no re-computation. Enough memory per GPU to support models with 128K+ token context windows.
Dedicated Blackwell Ultra clusters, designed and managed by CUDO
Deployed across 16 ISO-certified data centres
Scale to 1,000+ GPUs across dedicated multi-node clusters
High-speed networking with NVIDIA InfiniBand or Spectrum-X Ethernet. Verify per-GPU bandwidth with engineering
Expert rack-level design, installation, and benchmarking before handoff
Compatible with Slurm, Kubernetes, and NVIDIA Base Command
24/7 monitoring, management, and engineering support
Available at the most cost-effective pricing
Launch your AI products faster with on-demand GPUs and a global network of data center partners
Bare metal
Powered by renewable energy
No noisy neighbors
SpectrumX local networking
300Gbps external connectivity
NVMe SSD storage
Enterprise
Powerful GPU clusters
Scalable data center colocation
Large quantities of GPUs and hardware
Optimize to your requirements
Expert installation
Scale as your demand grows
Specifications
NVIDIA HGX B300 specifications
Starting from
Contact us for pricing
Architecture
NVIDIA Blackwell
GPU
8x NVIDIA Blackwell Ultra SXM
GPU memory
2.1 TB total, 62 TB/s HBM3e bandwidth
FP4 tensor core performance
144 PFLOPS | 108 PFLOPS
FP8 tensor core performance
72 PFLOPS
NVIDIA NVSwitch
2x
NVIDIA NVLink bandwidth
14.4 TB/s aggregate bandwidth
System power usage
~14.5 kW max
CPU
Intel Xeon 6776P processors
System memory
2 TB, configurable to 4 TB
Networking
8x OSFP ports serving 8x NVIDIA ConnectX-8 VPI. Up to 800 Gb/s of NVIDIA InfiniBand/Ethernet 2x dual-port QSFP112 NVIDIA BlueField®-3 DPU. Up to 400 Gb/s of NVIDIA InfiniBand/Ethernet
Management network
1GbE onboard network interface card (NIC) with RJ45 1GbE RJ45 host baseboard management controller (BMC)
Storage
OS: 2x 1.9 TB NVMe M.2, internal storage: 8x 3.84 TB NVMe U.2
Software
NVIDIA AI Enterprise (optimized AI software), NVIDIA Mission Control (AI data center operations and orchestration with NVIDIA Run:ai technology), NVIDIA DGX OS (operating system), supports Red Hat Enterprise Linux / Rocky / Ubuntu
Rack units (RU)
10
Operating temperature
10-35°C / 50-90°F
Where B300 clusters deliver the biggest impact
Explore uses cases for the NVIDIA B300 including Frontier model training, AI reasoning, large-scale inference, Sovereign and regulated AI.
AI reasoning and test-time scaling
Up to 288 GB HBM3e holds full model states in memory while Blackwell Ultra's enhanced Transformer Engine doubles attention throughput. The result is that reasoning models process longer chains of thought without offloading, maintaining output quality at significantly higher speed. An 8-GPU Blackwell node delivers over 250 tokens per second per user on DeepSeek R1-671B. All CUDO clusters are benchmarked and validated before handoff.
Large-scale inference at lower cost per token
Up to 5X higher throughput per megawatt compared to Hopper. More memory per GPU means fewer nodes to serve the same model. Lower infrastructure cost and less operational complexity.
Frontier model training and fine-tuning
Up to 1.5X faster training than B200 with FP4 precision across multi-node B300 clusters. Up to 2.3 TB of HBM3e per 8-GPU node holds larger model states in memory, reducing the need for complex model parallelism and the communication overhead it creates.
Sovereign and regulated AI
Deploy B300 clusters in ISO-certified data centres globally. Meet data residency and regulatory requirements with dedicated infrastructure under your control. Your hardware, your jurisdiction, managed by CUDO.
Blog
Resources
- Emmanuel Ohiri
Resources
- Emmanuel Ohiri
Resources
- Emmanuel Ohiri
Resources
- Emmanuel Ohiri
Resources
- Emmanuel Ohiri
Resources
- Emmanuel Ohiri
Browse alternative GPU solutions for your workloads
Access a wide range of performant NVIDIA and AMD GPUs to accelerate your AI, ML & HPC workloads
NVIDIA H100 PCIe
Price on request
Scale with high performance H100 GPUs on our reserved cloud.
