High-performance Bare Metal GPU

A complete NVIDIA HGX B300 node, in bare metal
8 Blackwell Ultra GPUs, 2.1 TB of HBM3e and NVLink 5, with no hypervisor or shared resources. Root access, choice of OS and CUDA driver version: you control the entire stack to train, fine-tune and serve your largest models.
Contact us
12-month minimum commitment


8 Blackwell Ultra GPUs: what has changed compared to the B200
Compared to the B200, the B300 features nearly 50% more HBM3e, delivers 1.5 times more dense FP4 compute and processes attention layers twice as fast. A trillion-parameter model in FP8 therefore fits into a single node, leaving nearly 1 TB for the KV cache, and the long-context reasoning models gain token throughput. The 8 GPUs form a single NVLink 5 domain at 1.8 TBps per GPU: tensor and expert parallelism do not leave the node.
2.1 TBof HBM3e memory on the node
108 PFLOPSof dense FP4 compute on the node
Benchmark
The B300 compared to benchmark GPUs
Choose a GPU: each card shows the difference compared to the B300.
AI computingDense FP8, PFLOPS
GPU memoryGB
Memory bandwidth:TBps
- Per HGR-AI-B300 server8 B300 GPUS
- 36 FP8 PFLOPS
- 2.1 TB of HBM3e
- 64 TBps of bandwidth
NVIDIA manufacturer data, per GPU, peak performance without sparsity. H100 and H200 in SXM version, RTX PRO 6000 Blackwell Server Edition (estimated dense FP8).
Technical specs
The HGR-AI-B300 in detail
| Component | Feature | HGR-AI-B300 |
|---|---|---|
| GPU8 × NVIDIA B300 | Model | 8 × NVIDIA HGX B300 SXM6Blackwell Ultra architecture |
| GPU memory | 2.1 TB HBM3e262 GB per GPU | |
| Memory bandwidth: | 64 TBps8 TBps per GPU | |
| FP4 compute (dense) | 108 PFLOPS13.5 PFLOPS per GPU | |
| FP8 compute (dense) | 36 PFLOPS4.5 PFLOPS per GPU | |
| BF16 compute (dense) | 18 PFLOPS2.25 PFLOPS per GPU | |
| GPU interconnection | 5th generation NVLink1.8 TBps per GPU, 14.4 TBps total | |
| CPU2 × Intel Xeon 6 | Processors | 2 × Intel Xeon 6776PIntel Xeon 6, Granite Rapids |
| Cores/Threads | 128 cores/256 threads2 × 64 cores | |
| Frequency | 2.3 GHz/3.9 GHzBase/turbo; 3.6 GHz all cores, up to 4.6 GHz with Priority Core Turbo (8 cores) | |
| L3 cache | 336 MB | |
| TDP | 350 W per CPUNo thermal throttling | |
| RAM2 TB DDR5 | Capacity | 2 TB DDR5 ECC |
| Frequency | 6400 MHz | |
| Storage38.4 TB of data | System | 3 × 3.84 TB PCIe 5.0 NVMe |
| Data | 5 × 7.68 TB PCIe 5.0 NVMeE1.S format, 38.4 TB total | |
| Network & bandwidthUp to 200 Gbps private | Public bandwidth | 10 Gbps guaranteed, unlimited15, 20 or 25 Gbps options |
| Private bandwidth | 100 Gbps guaranteed, unlimited2 × 100GbE NVIDIA ConnectX-6 Dx interfaces | |
| Link Aggregation | Up to 200 Gbps privateOVHcloud Link Aggregation on the 2 interfaces |
GPU: NVIDIA manufacturer data, peak performance without sparsity. Fixed hardware configuration; only the public bandwidth is adjustable.
Speak to an expert about the HGR-AI-B300
Leave your contact details: our team will get back to you to discuss your project.
Your questions answered
What uses is the HGR-AI-B300 intended for?
The HGR-AI-B300 is aimed at training and inference for large models, particularly reasoning models and long contexts. Its 8 interconnected GPUs and 2 TB of system memory are designed for long and intensive workloads. It is not aimed at double-precision scientific computing: the B300 prioritises FP4 and FP8, and its FP64 throughput is reduced.
Can I modify the server configuration?
The hardware configuration is fixed: 2 TB of DDR5, 8 NVIDIA B300 GPUs and the NVMe storage described above. Only the public bandwidth is adjustable, from 10 Gbps (included) to 25 Gbps.
In which datacentres is the HGR-AI-B300 available?
The HGR-AI-B300 is deployed in OVHcloud datacentres. Your sales contact will tell you the available locations for your project.
How does OVHcloud Link Aggregation work?
OVHcloud Link Aggregation groups the server’s 2 private 100GbE interfaces into a single logical link, meaning that you achieve 200 Gbps of aggregated private bandwidth, and benefit from redundancy in the event of a link failure.
What commitment periods are available?
The HGR-AI-B300 is offered with a minimum 12-month commitment. Online, you can select a 12 or 24-month commitment with an upfront payment. With an OVHcloud sales representative, 12, 24 or 36-month commitments with monthly payments are also possible. Plus, setup is free.