Einsteinium GPU Cluster¶
Einsteinium is an institutional GPU cluster that was deployed to meet the growing computational demand for researchers doing machine learning and deep learning. The system is named after the chemical element with symbol Es and atomic number 99 which was discovered at Lawrence Berkeley National Laboratory in 1952 and in honor of Albert Einstein who developed the theory of relativity.
es3 Partition¶
es3 is a partition consisting of RTX PRO 6000 Blackwell GPU nodes built on AMD EPYC processors, providing a high core count per node alongside four GPU cards. es3 partition nodes were introduced in 2026 to support growing AI/ML workflow needs. Five of the es3 nodes are condo nodes.
| Accelerator | Nodes | GPUs per Node/GPU Memory | CPU Processor | CPU Cores per Node | CPU RAM per Node | Infiniband |
|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell | 12 | 4x 96 GB | AMD EPYC 9754 | 128 | 384 GB or 768 GB | HDR |
How to specify desired GPU card(s)¶
Two qos values are available on the es3 partition:
| QoS | Limits | Notes |
|---|---|---|
es3_normal |
3-day max wall time, up to 5 nodes per user | Standard priority |
es3_lowprio |
Low priority (available to condo and ac_ accounts) |
Free of charge; jobs can be preempted and requeued when condo owners need their nodes. Not available pc_ accounts |
Checkpoint your es3_lowprio jobs
Jobs under es3_lowprio can be preempted at any time. Save checkpoint files regularly so that a preempted job can resume from its last checkpoint after it is requeued.
Due to hardware configuation, special attention is needed to ensure the ratio of CPU-core# to GPU#. You must request 32 CPU cores for each GPU card on the es3 partition. Together with the charging factor of 1.0 for the es3 partition, this means that each GPU-hour under es3_normal incurs at least 32 Service Units (SUs).
Examples:
- Request one RTX PRO 6000 card:
--partition=es3 --gres=gpu:RTX6000:1 --cpus-per-task=32 --ntasks=1 - Request four RTX PRO 6000 cards:
--partition=es3 --gres=gpu:RTX6000:4 --cpus-per-task=32 --ntasks=4
Example slurm script on es3
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --account=account_name
#SBATCH --partition=es3
#SBATCH --qos=es3_normal
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:RTX6000:1
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --account=account_name
#SBATCH --partition=es3
#SBATCH --qos=es3_normal
#SBATCH --nodes=1
#SBATCH --ntasks=4
#SBATCH --cpus-per-task=32
#SBATCH --gres=gpu:RTX6000:4
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
es2 Partition¶
es2 is a partition consisting of H100 and H200 GPU nodes. These nodes have been added in 2024 and 2025 to meet AI/LLM related research needs in addition to the growing use of GPUs in scientific computing.
| Accelerator | Nodes | GPUs per Node/GPU Memory | CPU Processor | CPU Cores per Node | CPU RAM per Node | Infiniband |
|---|---|---|---|---|---|---|
| NVIDIA H200 | 3 | 8x 141 GB | Intel Xeon Platinum 8570 | 112 | 2 TB | NDR |
| NVIDIA H100 | 4 | 8x 80 GB | Intel Xeon Platinum 8480+ | 112 | 1 TB | NDR |
How to specify desired GPU card(s)¶
The normal qos for the es2 partition is called es2_normal. The other qos values are es_debug and es_lowprio. Note that es_lowprio is not available to PI Computing Allowance (PCA) accounts (pc_*).
Due to hardware configuation, special attention is needed to ensure the ratio of CPU-core# to GPU#. You must request 14 CPU cores for each GPU card on the es2 partitions. Together with the charging factor of 4.0 for the es2 partition, this means that each GPU-hour on the es2 partition incurs at least 56 Service Units (SUs).
Examples:
- Request three H100 cards:
--partition=es2 --gres=gpu:H100:3 --cpus-per-task=14 --ntasks=3 - Request one H200 cards:
--partition=es2 --gres=gpu:H200:1 --cpus-per-task=14 --ntasks=1 --mem-per-cpu=18400M
As seen in the examples, use --mem-per-cpu=18400M for H200 nodes since they have more memory than the default on the es2 partition.
Example slurm script on es2
Here is an example slurm script that requests three NVIDIA H100 GPU cards.
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --account=account_name
#SBATCH --partition=es2
#SBATCH --qos=es2_normal
#SBATCH --nodes=1
#SBATCH --ntasks=3
#SBATCH --cpus-per-task=14
#SBATCH --gres=gpu:H100:3
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --account=account_name
#SBATCH --partition=es2
#SBATCH --qos=es2_normal
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=14
#SBATCH --gres=gpu:H200:1
#SBATCH --mem-per-cpu=18400M
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
es1 Partition¶
es1 is a partition consisting of multiple GPU node types to address the different research needs. These include:
| Accelerator | Nodes | GPUs per Node/GPU Memory | CPU Processor | CPU Cores per Node | CPU RAM per Node | Infiniband |
|---|---|---|---|---|---|---|
| NVIDIA A100 | 1 | 4x 80 GB | AMD EPYC 7713 | 64 | 512 GB | HDR |
| NVIDIA A40 | 30 | 4x 48 GB | AMD EPYC 7742 | 64 | 512 GB | FDR |
| NVIDIA GRTX8000 | 1 | 4x 48 GB | AMD EPYC 7713 | 64 | 512 GB | HDR |
| NVIDIA V100 | 15 | 2x 32 GB | Intel Xeon E5-2623 | 8 | 64GB or 192GB | FDR |
How to specify desired GPU card(s)¶
The normal qos for the es1 partition is called es_normal.
Due to hardware configuation, special attention is needed to ensure the ratio of CPU-core# to GPU#.
Examples:
- Request one V100 card:
--cpus-per-task=4 --gres=gpu:V100:1 --ntasks=1 - Request two A40 cards:
--cpus-per-task=16 --gres=gpu:A40:2 --ntasks=2 - Request one A100 cards:
--cpus-per-task=16 --gres=gpu:A100:1 --ntasks=1 - Request four GRTX8000 cards:
--cpus-per-task=16 --gres==gpu:GRTX8000:4 --ntasks=4
Example slurm script on es1
Here is an example slurm script that requests one NVIDIA A40 GPU card.
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --account=account_name
#SBATCH --partition=es1
#SBATCH --qos=es_normal
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --gres=gpu:A40:1
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
#!/bin/bash
#SBATCH --job-name=test
#SBATCH --account=account_name
#SBATCH --partition=es1
#SBATCH --qos=es_normal
#SBATCH --nodes=1
#SBATCH --ntasks=2
#SBATCH --cpus-per-task=4
#SBATCH --gres=gpu:V100:2
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
es0 Partition¶
es0 is a partition with NVIDIA 2080 TI GPUs that do not incur Service Unit (SU) charges.
| Accelerator | Nodes | GPUs per Node/GPU Memory | CPU Processor | CPU Cores | CPU RAM | Infiniband |
|---|---|---|---|---|---|---|
| NVIDIA 2080TI | 12 | 4x 11 GB | Intel Xeon Silver 4212 | 8 | 96GB | FDR |
Example slurm script on es0
#!/bin/bash
#SBATCH --job-name=testes0
#SBATCH --account=account_name
#SBATCH --partition=es0
#SBATCH --qos=es_normal
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=2
#SBATCH --gres=gpu:1
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py
#!/bin/bash
#SBATCH --job-name=testes0
#SBATCH --account=account_name
#SBATCH --partition=es0
#SBATCH --qos=es_normal
#SBATCH --nodes=1
#SBATCH --ntasks=4
#SBATCH --cpus-per-task=2
#SBATCH --gres=gpu:4
#SBATCH --time=1:00:00
module load ml/pytorch
python train.py