DGX Cloud Benchmarking - Performance Recipes

Performance Recipes are ready-to-use templates for evaluating performance of specific AI use cases across hardware and software combinations. These containerized recipes allow users to quickly set up and run standardized benchmarking methodology in their own environment, ensuring consistent and comparable results across platforms.

These Performance Recipes support performance characterization

• across a variety of defined AI workloads, including pre-training, fine tuning, and inference.
• across GPU-based infrastructure, whether running on-premises or with cloud service providers (CSPs).

Each recipe maps to one workload and can be run at various cluster scales and precisions. These workloads are tested against the NVIDIA Reference Architecture and those results are provided as a baseline for comparison. These performance metrics are collected from production environments and are subject to real-world variability.

Prerequisites

To use the Performance Recipes, make sure you have the following prerequisites installed on your cluster:

General Prerequisites

• Bash 4.2 or newer
• NGC Registry Access
• NGC CLI 3.148.1 or newer (Optional, required for NIM Inference workloads)
• Python 3.12.x
• CUDA 12.3 or newer
• NV Driver: 535.129.03 or newer
• OFED: 5.9-0.5.6.0.127 or newer
• NCCL: 2.19.4 or newer

Cluster-Specific Prerequisites

Depending on your cluster's job scheduler, ensure the following are met:

• Slurm Clusters
  • Version 22.x or newer
  • task/affinity plugin required for process pinning
  • Enroot

Quick Start Guide

1. Clone the repository:

   git clone https://github.com/NVIDIA/dgxc-benchmarking.git
   cd dgxc-benchmarking

2. (Optional) For NIM Inference workloads only:

   • Generate an NGC API key from the NGC Registry
   • Install and configure the NGC CLI:
   x86

   curl -L https://ngc.nvidia.com/downloads/ngccli_linux.zip -o ngccli_linux.zip
   unzip -q ngccli_linux.zip -d $HOME/.local/bin
   rm ngccli_linux.zip
   export PATH=$HOME/.local/bin:$PATH
   ngc config set

   arm64

   curl -L https://ngc.nvidia.com/downloads/ngccli_arm64.zip -o ngccli_arm64.zip
   unzip -q ngccli_arm64.zip -d $HOME/.local/bin
   rm ngccli_arm64.zip
   export PATH=$HOME/.local/bin/ngc-cli:$PATH
   ngc config set

3. Run the installer:

   # The installer will add packages to your current Python environment.
   # We recommend to activate a virtual environment with python 3.12.x before running installer command below.
   
   ./install.sh

   The installer will:

   • Install required Python packages in your current environment
   • Set up your benchmarking environment
   • Configure SLURM settings
   • Let you select workloads to install
   • Prepare all necessary dependencies

   Note: For detailed installation options, workload-specific virtual environments, and troubleshooting, see the Installer README.

4. Run a benchmark:

   # Navigate to your installed workload directory
   cd $LLMB_INSTALL
   
   # Example: Run Nemotron4 340B pretraining on 256 GPUs with FP8 precision
   llmb-run single -w pretraining_nemotron -s 340b --dtype fp8 --scale 256

Directory Layout and Key Variables

After running the installer, the following directory structure and environment variables are used:

• LLMB_REPO: Directory containing the clone of the recipe repository.
• LLMB_INSTALL: Top-level directory for all benchmarking artifacts (images, datasets, venvs, workloads, etc).
• LLMB_WORKLOAD: Workload-specific directory, e.g. ${LLMB_INSTALL}/workloads/pretraining_nemotron.
• Results, logs, and checkpoints are stored under subfolders of LLMB_WORKLOAD (see below).

Example structure:

$LLMB_INSTALL/
  ├── images/
  ├── datasets/
  ├── venvs/
  └── workloads/
        └── pretraining_nemotron/   # <- $LLMB_WORKLOAD
              ├── NeMo/
              ├── ...
              └── experiments/

Migration Note: If you previously used STAGE_PATH, replace it with LLMB_INSTALL (top-level) and LLMB_WORKLOAD (workload-specific). All output, logs, and checkpoints are now under LLMB_WORKLOAD.

Workload Resources Overview

Each workload resource includes:

• Configuration details: Comprehensive software and hardware setup information.
• Performance scripts: Predefined scripts to generate and analyze performance results.

The overview page for each workload highlights target performance metrics for the specified configuration, focusing on speed measurements such as the time taken per training step and the number of tokens processed per second.

Available Benchmarks

The following tables list each benchmark used to evaluate the model's performance, along with their specific configurations.

Note: The "Scale (# of GPUs)" column indicates the minimum supported scale and the maximum scale tested for each workload. The recipes may function at larger scales, although they have not been explicitly validated beyond the listed maximum.

GB200 Workloads

Framework	Model	Container Version	Model Size	Type	Scale (# of GPUs)	Precision	Model Access Required	Checkpointing	Cluster Type
NeMo	Nemotron4	25.04.00	340B	Pretrain	128-256	FP8, BF16	No	No	Slurm
NeMo	Llama 3.1	25.04.01	405B	Pretrain	128-256	FP8, BF16	Yes	No	Slurm
NeMo	DeepSeek V3	25.04.01	671B	Pretrain	128-256	BF16	Yes	No	Slurm
NeMo	Grok1	25.04.00	314B	Pretrain	128-256	FP8, BF16	Yes	No	Slurm
NeMo	Llama4 Maverick	25.04.01	400B	Pretrain	128-512	FP8, BF16	Yes	No	Slurm

H100 Workloads

Baseline performance metrics were using workloads on the NVIDIA DGX H100 Reference Architecture. For more information see DGX H100 Systems.

Framework	Model	FW Container Version	Model Size	Type	Scale (# of GPUs)	Precision	Model Access Required	Checkpointing	Cluster Type
NeMo	Nemotron4	25.04.00	15B	Pretrain	16	FP8, BF16	No	Yes	Slurm
NeMo	Nemotron4	25.04.00	340B	Pretrain	256	FP8, BF16	No	Yes	Slurm
NeMo	DeepSeek V3	25.04.01	671B	Pretrain	1024	BF16	Yes	No	Slurm
NeMo	Llama4 Maverick	25.04.01	400B	Pretrain	512	FP8, BF16	Yes	No	Slurm
NeMo	Nemotron-H	25.04.01	56B	Pretrain	32-2048	FP8	No	No	Slurm
NeMo	Llama 3.1	25.04.01	405B	Pretrain	512	FP8, BF16	Yes	No	Slurm
NeMo	Grok1	25.04.00	314B	Pretrain	512	FP8, BF16	Yes	No	Slurm
NIM	Llama 3.1 and 3.2	instruct:1.3.3, rerank:1.3, embed:1.3.1	70b, 1b	Inference	1-8	n/a	Yes	No	Slurm
NIM	Llama 3	1.0.3	70B	Inference	4	FP8	Yes	No	Slurm
NIM	DeepSeek R1	1.7.2	671B	Inference	16	FP8	Yes	No	Slurm

Deprecated

Framework	Model	FW Container Version	Model Size	Type	Scale (# of GPUs)	Precision	Model Access Required	Checkpointing	Cluster Type	Last Version
Maxtext	Llama3	25.01	70B	Pretrain	128-2048	FP8, BF16	No	No	Slurm	25.04.01
NeMo	Llama 3	24.12	8B, 70B	Fine-Tuning (SFT, LORA)	8-32	FP8, BF16	Yes	No	Slurm	25.04.01
NeMo	GPT3	24.12	175B	Pretrain	128-2048	FP8, BF16	No	No	Slurm	25.04.01
HF	Llama 2	24.02-py3	70B	Finetuning	8-512	FP8, BF16	Yes	No	Slurm	25.01.01
HF	Mistral	24.02-py3	7B	Finetuning	8-256	FP8, BF16	Yes	No	Slurm	25.01.01
Jax	Llama 2	jax:maxtext-2024-12-09	70B	Pretrain	128-2048	FP8, BF16	No	No	Slurm	25.01.01
Jax	GPT3	jax:pax-2024-03-04	175B	Pretrain	128-2048	FP8, BF16	No	No	Slurm	25.01.01
NeMo	Llama 2	24.03.01.framework	7B	Pretrain	8-2048	FP8, BF16	Yes	No	Slurm	25.08
NeMo	Llama 2	24.03.01.framework	70B	Pretrain	64-2048	FP8, BF16	Yes	No	Slurm	25.08
NeMo	Llama 3	25.05	8B	Pretrain	8-128	FP8, BF16	Yes	No	Slurm	25.08
NeMo	Llama 3	25.05	70B	Pretrain	64-2048	FP8, BF16	Yes	No	Slurm	25.08

Reference Infrastructure

The LLM Benchmarking Collection published baseline benchmark results using the following reference infrastructures, CSP-specific configurations, and software.

H100 Reference Architecture

Baseline performance metrics for H100 workloads were collected using the NVIDIA DGX H100 Reference Architecture. For more information see DGX H100 Systems.

• GPU: 8xH100 80GB HBM3 (640GB total)
  • TDP 700W
  • Memory bandwidth 3.2 TB/s
• CPU: 2x Intel Sapphire Rapids, Intel(R) Xeon(R) Platinum 8480CL E5
  • 112 cores (56 cores per CPU)
  • 2.00 GHz (Base), 3.8 GHz (Max boost)
  • Numa nodes per socket = 1
  • PCIe Gen5
• NVLink: NVLink 4th Generation
  • 900 GB/s (GPU to GPU NVLink bidirectional bandwidth)
  • 18 Links per GPU
• InfiniBand:
  • Compute links: 8x 400 Gbit/s
  • Storage links: 2x 400 Gbit/s
• System Memory: 2TB
• Local Storage:
  • 2x 1.92TB NVMe M.2
  • 8x 3.84TB NVMe U.2

GB200 Reference Architecture

Baseline performance metrics for GB200 workloads were collected using the NVIDIA DGX GB200 Reference Architercture. For more information see NVIDIA GB200 NVL72

• GB200 Grace Blackwell Superchip
  • 72 Arm Neoverse V2 cores
  • 2x Blackwell GPUs
    • Memory bandwidth 16 TB/s
• NVLink: NVLink 5th Generation
  • 3.6 TB/s

CSP Specific Configurations

AI platforms may vary in implementation, such as differences in network fabric and virtualization implementations, and thus require different tuning. For optimal performance, users should leverage the correct implementation for their platform. The example platform-specific tuning is provided as a starting point. Further tuning may be necessary if instance type varies from the Reference Architecture.

AWS

For NeMo based images EFA support is already included starting with version 25.02 (nvcr.io/nvidia/nemo:25.02).

For other images or if you need to update Enable Elastic Fabric Adapter (EFA) follow the step-by-step guide. Use the reference NCCL tests Dockerfile with EFA support.

GCP

Ensure that all required pre-conditions for GCP cluster deployment have been met.

Configure Compute Fabric with TCP-X by ensuring the following environment variables are set and present for your environment.

NCCL_LIB_DIR='/var/lib/tcpxo/lib64' source /var/lib/tcpxo/lib64/nccl-env-profile.sh; \
	  export NCCL_FASTRAK_CTRL_DEV=enp0s12; \
	  export NCCL_FASTRAK_IFNAME=enp6s0,enp7s0,enp13s0,enp14s0,enp134s0,enp135s0,enp141s0,enp142s0; \
	  export NCCL_SOCKET_IFNAME=enp0s12; \
	  export NCCL_FASTRAK_LLCM_DEVICE_DIRECTORY=/dev/aperture_devices; \
	  export NCCL_NET=FasTrak; \
	  ls /var/lib/tcpxo/lib64;"

Important:

• The above example hasn't been tested with the latest TCP-X version. Check with your cluster admin for the most recent instructions.
• If additional files need to be mounted into running container, they should be placed under $LLMB_WORKLOAD folder as this location is already mounted.

Azure

Requires two settings for optimal performance:

1. NCCL_TOPO_FILE=<path to topo file under $LLMB_WORKLOAD>.
   • The VM topology files ensure that the correct CPUs, GPUs and NICs are bound together. Location of this file varies, it must be mounted into the container.
   • Important: Place NCCL Topology file under $LLMB_WORKLOAD folder as this location is already mounted into running container.
2. NCCL_P2P_CHUNKSIZE=2097152
   • Increasing message size for NCCL send/recv for optimal performance

Example Configuration for a training recipe:

export NCCL_TOPO_FILE=$LLMB_WORKLOAD/nvd5-topo.xml # Exact location varies by cluster
export NCCL_P2P_NET_CHUNKSIZE=2097152

Release Notes

Performance Recipes version 25.05.01

Added

• H100 support for the following recipes
  • Llama3.1 405B
  • Grok1 314B

Changed

• Fixed DeepSeek and Llama4 READMEs
• Installer
  • Enforce min and max python versions
  • Use GRES properly for CPU_PARTITION jobs.
• bitsandbytes update for ARM systems.

FAQ

Contains synopsis and resolution for known issues

1. Training logs contain multiple userbuffers.cu messages

Symptom

Large scale pre-training run logs contain message like below:

[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 18 [2]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 18 [4]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 19 [2]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 19 [4]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 22 [2]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 22 [4]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 23 [2]: expecting 1 got 0
[userbuffers.cu:userbuffers_fp16_sum_inplace_gpu_rr_rs_oop_fp8:797] [6] Reduce-scatter: SM 23 [4]: expecting 1 got 0

Solution

These usually mean that one of the GPUs is hanging. Possible resolutions:

• re-running the job on a different set of nodes
• rebooting affected nodes.

2. Slurm job failed, need to find log files

Symptom

A Slurm job failed during benchmark run. E.g., a nemotron benchmark job with ID=2041792 failed

sacct -j 2041792
JobID           JobName  Partition    Account  AllocCPUS      State ExitCode
------------ ---------- ---------- ---------- ---------- ---------- --------
2041792        launch.sh     batch test              224     FAILED      1:0
2041792.bat+      batch            test              224     FAILED      1:0
2041792.ext+     extern            test              224  COMPLETED      0:0
2041792.0          bash            test              224     FAILED      1:0

Solution

NeMo2 (e.g., Nemotron4, Llama3.1)

You can find log files associated with this run under $LLMB_WORKLOAD/experiments/pretrain_nemotron4_<size>_<dtype>_<scale>_<config> folder. The folder will have subfolders that will contain log-account.pretrain_nemotron4_<size>_<dtype>_<scale>_<config>.out files with a root cause error message.

E.g., for the job failure above and assuming the nemotron 15b job ran on 16 GPUs, used version 25.05, and with precision bf16 the path will be under $LLMB_WORKLOAD/experiments/pretrain_nemotron4_15b_bf16_gpus16_tp1_pp1_cp1_vp1_mbs2_gbs64/...

Search for errors in the log-account.pretrain_nemotron4_15b_bf16_gpus16_tp1_pp1_cp1_vp1_mbs2_gbs64_3358926_0.out file.

3. Unable to use venv required by benchmark

Symptom

If a benchmark requires virtual python environment (venv) but virtualenv executable isn't available on the login node and/or login nodes cannot be updated by non-sudo users, you would see errors like below when trying to setup venv

bash-5.2$ virtualenv
bash: virtualenv: command not found

Solution

There are alternative virtual environment options available like conda.

To install and activate conda virtual environment

# pick INSTALL_PATH with sufficient disk space
INSTALL_PATH=~
wget -q https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O $INSTALL_PATH/miniconda.sh
bash $INSTALL_PATH/miniconda.sh -b -p $INSTALL_PATH/miniconda3
$INSTALL_PATH/miniconda3/bin/conda init
source ~/.bashrc

When you are finished running this benchmark you can deactivate the environment, run this command

conda deactivate

Support

Terminology used in these recipes is explained in the Appendix.

For questions or to provide feedback, please contact [email protected]