Oracle Cloud Infrastructure Documentation

OCI GPU Quick Start: NVIDIA B300

This document provides hardware specifications, supported OS images, onboarding verification, sample benchmarks, and best-practices for OCI deployments using the NVIDIA B300 GPU shape.

BM.GPU.B300.8 is a high-density GPU bare-metal shape built around eight NVIDIA B300 SXM6 AC GPUs, dual Intel Xeon Platinum 8592+ processors, 4 TB of DDR5 system memory, and high-bandwidth RoCE networking for both single-node and scale-out AI workloads.

At a Glance

  • Shape: BM.GPU.B300.8
  • GPU configuration: 8 x NVIDIA B300 SXM6 AC
  • Recommended OS baseline: Oracle Linux 9+ or Ubuntu Linux 24.04+
  • Recommended software baseline: NVIDIA Driver 590+ (Open), CUDA 13.1+, NCCL 2.28.9+
  • Primary verification command: nvidia-smi
  • Operational profile: single-node and scale-out AI workloads

When To Use This Shape

Use this shape when you need top-end single-node NVIDIA performance with a current Blackwell software stack and a clear path to scale-out training.

It is best suited for customers validating new-generation NVIDIA performance, building image baselines for B300, or moving from smoke tests into NCCL and model-training checks quickly.

First 15 Minutes

  1. Launch the shape with an approved image from the Provided Images table.
  2. Run nvidia-smi and confirm all eight B300 GPUs and the expected driver stack are visible.
  3. Follow the Hello World Verification container smoke test.
  4. Run the first all-reduce benchmark and confirm the expected large-message bandwidth range before exploring the rest of the benchmark matrix.

Table of Contents

Hardware Specifications

Shape Name GPU Model GPUs/Node GPU Memory (GB/GPU) GPU Memory Total CPU # of CPUs System Memory Local Storage Host NIC RDMA (ROCe) NICs
BM.GPU.B300.8 B300 8 288 2.1 TB 2 x Intel Xeon Platinum 8592+ @ 1.9 GHz 128 Cores (256 with HT) 4 TB DDR5 8 x 3.5 TB NVMe = 28 TB 100 Gb/s 8 x 2 x 400 Gb/s = 6.4 Tb/s

See the OCI Compute Shapes Docs for up-to-date details.

Recommended Operating Systems

  • Oracle Linux 9+
  • Ubuntu Linux 24.04+
  • DOCA OFED 3.2.1+
  • NVIDIA Driver 590+ (Open)
  • CUDA 13.1+
  • NCCL 2.28.9+
  • HPC-X 2.25.1+
  • Oracle Cloud Agent 1.57.0+

Custom OS Image Creation with Packer

To build your images using packer clone the OCI HPC Images repo and run the commands found there OCI HPC Images GitHub Repo.

Provided Images

OS Version Image Packer Build Details OCI Platform Image Link Driver Versions
OCI GPU AI Image with Ubuntu Linux 22.04 Canonical-Ubuntu-22.04-DOCA-OFED-3.2.1-GPU-580-OPEN-CUDA-13.0 PAR Link NVIDIA OPEN 580, DOCA OFED 3.2.1, CUDA 13.0, OCA 1.57.0
OCI GPU AI Image with Ubuntu Linux 22.04 Canonical-Ubuntu-22.04-DOCA-OFED-3.2.1-GPU-590-OPEN-CUDA-13.1 PAR Link NVIDIA OPEN 590, DOCA OFED 3.2.1, CUDA 13.1, OCA 1.57.0
OCI GPU AI Image with Ubuntu Linux 24.04 Canonical-Ubuntu-24.04-6.8-DOCA-OFED-3.2.1-GPU-580-OPEN-CUDA-13.0 PAR Link NVIDIA OPEN 580, DOCA OFED 3.2.1, CUDA 13.0, Kernel 6.8, OCA 1.57.0
OCI GPU AI Image with Ubuntu Linux 24.04 Canonical-Ubuntu-24.04-6.8-DOCA-OFED-3.2.1-GPU-590-OPEN-CUDA-13.1 PAR Link NVIDIA OPEN 590, DOCA OFED 3.2.1, CUDA 13.1, Kernel 6.8, OCA 1.57.0
OCI GPU AI Image with Ubuntu Linux 24.04 Canonical-Ubuntu-24.04-6.14-DOCA-OFED-3.2.1-GPU-590-OPEN-CUDA-13.1 PAR Link NVIDIA OPEN 590, DOCA OFED 3.2.1, CUDA 13.1, Kernel 6.14, OCA 1.57.0
OCI GPU AI Image with Oracle Linux 8 Oracle-Linux-8.10-RHCK-DOCA-OFED-3.2.1-GPU-580-OPEN-CUDA-13.0 PAR Link NVIDIA OPEN 580, DOCA OFED 3.2.1, CUDA 13.0, OCA 1.57.0
OCI GPU AI Image with Oracle Linux 8 Oracle-Linux-8.10-RHCK-DOCA-OFED-3.2.1-GPU-590-OPEN-CUDA-13.1 PAR Link NVIDIA OPEN 590, DOCA OFED 3.2.1, CUDA 13.1, OCA 1.57.0
OCI GPU AI Image with Oracle Linux 9 Oracle-Linux-9.7-RHCK-DOCA-OFED-3.2.1-GPU-580-OPEN-CUDA-13.0 PAR Link NVIDIA OPEN 580, DOCA OFED 3.2.1, CUDA 13.0, OCA 1.57.0
OCI GPU AI Image with Oracle Linux 9 Oracle-Linux-9.7-RHCK-DOCA-OFED-3.2.1-GPU-590-OPEN-CUDA-13.1 PAR Link NVIDIA OPEN 590, DOCA OFED 3.2.1, CUDA 13.1, OCA 1.57.0

Hello World Verification

Run nvidia-smi to verify that all eight GPUs are visible and healthy:

nvidia-smi

Expected output:

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 590.48.01    Driver Version: 590.48.01    CUDA Version: 13.1               |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA B300 SXM6 AC           On  |   00000000:14:00.0 Off |                    0 |
|  N/A   34C    P0             238W / 1100W |     0MiB / 275040MiB |      0%      Default |
|                                         |                        |           Disabled   |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA B300 SXM6 AC           On  |   00000000:32:00.0 Off |                    0 |
...
|   7  NVIDIA B300 SXM6 AC           On  |   00000000:DD:00.0 Off |                    0 |
+-----------------------------------------------------------------------------------------+

You should see all eight GPUs listed with no ECC errors and expected idle power draw.

Performance Benchmarks

NVIDIA publishes NCCL as the primary collective communication library for multi-GPU AI and HPC workloads. The benchmark data collected for this shape includes both collective communication results and model-training performance, with larger message sizes being the most useful for cluster planning.

All Reduce - Single Node

. /opt/hpcx-v2.25.1-gcc-doca_ofed-ubuntu24.04-cuda13-x86_64/hpcx-init.sh
hpcx_load
/opt/oci-hpc/nccl-tests/build/all_reduce_perf -b 512k -e 8G -f 2 -g 8

# Collective test starting: all_reduce_perf
# nThread 1 nGpus 8 minBytes 524288 maxBytes 8589934592 step: 2(factor) warmup iters: 1 iters: 20 agg iters: 1 validation: 1 graph: 0
NCCL version 2.28.9+cuda13.1
#
#                                                              out-of-place                       in-place
#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)
# <smaller message sizes truncated>
 8589934592    2147483648     float     sum      -1    18006  477.05  834.84      0    17990  477.49  835.61      0
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 440.608
#
# Collective test concluded: all_reduce_perf

All Reduce - 2 Nodes

NCCL version 2.28.9+cuda13.1
#
#                                                              out-of-place                       in-place
#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)
# <smaller message sizes truncated>
 8589934592    2147483648     float     sum      -1    21012  408.82  766.54      0    21036  408.35  765.65      0
17179869184    4294967296     float     sum      -1    42007  408.98  766.84      0    41935  409.68  768.15      0
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 405.346
#
# Collective test concluded: all_reduce_perf

All-to-All - Single Node

# Collective test starting: alltoall_perf
# nThread 1 nGpus 8 minBytes 524288 maxBytes 8589934592 step: 2(factor) warmup iters: 1 iters: 20 agg iters: 1 validation: 1 graph: 0
NCCL version 2.28.9+cuda13.1
#
#                                                              out-of-place                       in-place
#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)
# <smaller message sizes truncated>
 8589934592   268435456     float    none      -1    11126  772.06  675.55      0    11206  766.58  670.76    N/A
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 359.21
#
# Collective test concluded: alltoall_perf

All-to-All - 2 Nodes

For guidance on running additional NCCL collective benchmarks on this GPU family, see the NCCL user guide.

# Collective test starting: alltoall_perf
# nThread 1 nGpus 1 minBytes 524288 maxBytes 8589934592 step: 2(factor) warmup iters: 1 iters: 50 agg iters: 1 validation: 1 graph: 0
#
#                                                              out-of-place                       in-place
#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)
# <smaller message sizes truncated>
 8589934592   134217728     float    none      -1    44498  193.04  180.98      0    44500  193.03  180.97    N/A
# Out of bounds values : 0 OK
# Avg bus bandwidth    : 102.495
#
# Collective test concluded: alltoall_perf

Model Inference Performance

Model-training benchmark results collected for this shape:

Model Scale (GPUs) Dtype MODEL_TFLOPS/GPU (Mean)
Qwen3 30B (MoE) 64 BF16 648.77
Qwen3 30B (MoE) 32 BF16 649.06
Qwen3 30B (MoE) 16 BF16 649.74
Qwen3 30B (MoE) 8 BF16 648.11
Llama 3.1 70B 64 FP8 1762.80
Nemotron-H 56B 64 FP8 1637.74
Nemotron-H 56B 32 FP8 1646.43

OKE GPU Getting Started

Information on getting up and running on OKE can be found here.

Useful B300-specific OKE starting points in oci-hpc-oke:

Troubleshooting

This guide includes a broad health-check set covering GPU visibility, firmware, NUMA topology, RDMA connectivity, storage, and intra-node bandwidth checks.

GPU Visibility

nvidia-smi

Use this to verify that GPUs 0-7 are present and healthy.

VBIOS Version

nvidia-smi -q | grep -i "vbios version"

Sample output:

VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05
VBIOS Version : 97.10.64.00.05

NUMA Layout

numactl --hardware

The expected topology is a two-node NUMA layout with roughly 2 TB of memory per NUMA node on a healthy system.

RDMA Interface State

rdma link

The front-end network is eth0 on mlx5_6. The RoCE RDMA interfaces should report state ACTIVE and physical_state LINK_UP.

Point-to-Point RDMA Health

ib_write_bw and ib_write_lat are standard inter-node health checks for this platform.

Healthy target ranges for this platform:

  • ib_write_bw: about 384+ Gbps
  • ib_write_lat: about ~3 usec for 2-8 byte messages on nodes connected to the same switch

Additional Validation Tools

Additional recommended validation tools:

  • gpu-fryer for sustained thermal and FLOPS checks
  • nvbandwidth for intra-node NVLink validation
  • BabelStream for memory-bandwidth checks
  • dcgmi diag -r [1,2,3,4] for layered GPU diagnostics

Example commands:

docker run --gpus all ghcr.io/huggingface/gpu-fryer:latest 60

git clone https://github.com/NVIDIA/nvbandwidth
cd nvbandwidth
sudo cmake -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc -DCMAKE_CUDA_ARCHITECTURES=native
sudo ./debian_install.sh
sudo chown -R $USER:$USER CMakeFiles
make
./nvbandwidth -t device_to_device_bidirectional_memcpy_read_ce

git clone https://github.com/UoB-HPC/BabelStream.git
cd BabelStream
git checkout v5.0
mkdir build
cmake -B build -DMODEL=cuda -DCMAKE_CUDA_COMPILER=/usr/local/cuda-13.1/bin/nvcc -DCUDA_ARCH=sm_100
cd build
make
./cuda-stream

dcgmi diag -r [1,2,3,4]

Further Reading & Support

Additional references:

The GPU thermal profile is not perfectly uniform across all eight devices because of board layout and airflow. If some GPUs consistently run warmer than others while staying within normal limits, that can still be expected behavior on this platform.