ACES

ACES (Accelerating Computing for Emerging Sciences) is a Texas A&M HPRC testbed built around composable infrastructure: ACES pools GPUs, FPGAs, and other accelerators separately from its CPU nodes and lets you assemble custom node configurations on demand over a high-speed PCIe fabric rather than fixing specific accelerators to specific nodes. This makes it a strong fit for researchers who want to experiment with accelerator combinations rather than commit to a single fixed hardware setup up front.

ACES is also notable for hosting accelerator types that are hard to find elsewhere, including Graphcore IPUs, NextSilicon co-processors, and NEC Vector Engines, alongside more familiar NVIDIA H100 and Intel GPUs/FPGAs. This breadth makes it especially useful for AI/ML researchers and others looking to benchmark or prototype on emerging hardware architectures before committing to a particular accelerator for production-scale work.

Jobs

Jobs on ACES are scheduled by Slurm and charged in Service Units (SUs), where 1 SU equals one core-hour, one core used for one hour of walltime. For example, using a full compute node (96 cores) for one hour costs 96 SUs. Jobs that use GPUs or other accelerators are charged at a higher rate on top of the core charge; current accelerator rates are listed on the HPRC Account Management System (AMS) page. SUs are drawn from the account named in the job, or your default account if none is specified.

Each job selects a partition (with --partition), which determines the hardware it runs on, and must request a core count and a walltime limit; if no walltime is given, the partition default applies. Compute nodes provide 96 cores and 512 GB of memory (488 GB usable for jobs), and walltime limits vary by partition. If you want more info about the compute nodes or the system architecture, you can check out the Hardware page. For more detail, please visit the Batch System page where the building job files basics are explained in depth.

The Batch System page also provides information about required batch job specifications and optional/alternative specifications. Here is an example on how to run a serial job (single core, single node):

#!/bin/bash

##NECESSARY JOB SPECIFICATIONS
#SBATCH --job-name=JobExample1       #Set the job name to "JobExample1"
#SBATCH --time=01:30:00              #Set the wall clock limit to 1hr and 30min
#SBATCH --ntasks=1                   #Request 1 task
#SBATCH --mem=5000M                  #Request 5000MB (5GB) per node
#SBATCH --output=Example1Out.%j      #Send stdout/err to "Example1Out.[jobID]"

##OPTIONAL JOB SPECIFICATIONS
##SBATCH --account=123456             #Set billing account to 123456
##SBATCH --mail-type=ALL              #Send email on all job events
##SBATCH --mail-user=email_address    #Send all emails to email_address

#First Executable Line

Note: The `pvc` node has 2–8 GPUs per node, varies by node.

The Grace-Hopper node (gh01) is not a batch partition; it is reached by connecting to it directly over SSH from an ACES login node. Its RAM varies 256–512+ GB.

Note on node counts: ACES uses a composable Liqid PCIe fabric, so accelerators (H100, A30, PVC, FPGA, NextSilicon) are dynamically attached to a shared pool of Sapphire Rapids compute nodes rather than being fixed to a partition. A single node can belong to several partitions at once, live gpuavail output shows nodes assigned to cpu,pvc simultaneously, so the per-queue node counts overlap and are not additive (they sum to more than the 110-node Sapphire Rapids pool by design). Counts are configured totals from live sinfo and include nodes that were offline/drained at capture time. For a real-time view, run sinfo -s -p <partition> or gpuavail on a login node.

Queue specifications

Metrics updated 2026-06-16

Name Purpose Nodes CPU cores / node GPUs / node Node RAM Jobs
30 days
Wait Time
30-day trend
Wall Time
30-day trend
cpu General CPU-only jobs 88 Intel Xeon 8468 (Sapphire Rapids) (96 cores) 488 GB 91,196
cpu wait time: average 1.2 hours, range 0 to 5.4 hours over 30 days
cpu wall time: average 1.4 hours, range 0.7 to 4.3 hours over 30 days, wall-time limit 72h
gpu NVIDIA GPU workloads (AI/ML, CUDA, parallel GPU jobs). High-memory GPU nodes 8 Intel Xeon 8468 (Sapphire Rapids) (96 cores) 2 NVIDIA H100 488 GB 2,621
gpu wait time: average 7.3 hours, range 0 to 32.2 hours over 30 days
gpu wall time: average 3.4 hours, range 0.8 to 6.7 hours over 30 days, wall-time limit 48h
gpu_debug Short GPU testing/debugging (1 node max, 2-hour limit). Mixes A30s and H100s across nodes. 3 Intel Xeon 8468 (Sapphire Rapids) (96 cores) 2 NVIDIA A30 488 GB 628
gpu_debug wait time: average 2.2 hours, range 0 to 15.4 hours over 30 days
gpu_debug wall time: average 0.9 hours, range 0 to 1.7 hours over 30 days, wall-time limit 2h
gpu_debug Short GPU testing/debugging (1 node max, 2-hour limit). Mixes A30s and H100s across nodes. 3 Intel Xeon 8468 (Sapphire Rapids) (96 cores) 2 NVIDIA H100 488 GB 628
gpu_debug wait time: average 2.2 hours, range 0 to 15.4 hours over 30 days
gpu_debug wall time: average 0.9 hours, range 0 to 1.7 hours over 30 days, wall-time limit 2h
pvc Intel GPU Max (PVC) jobs — AI/ML, LLM inference 30 Intel Xeon 8468 (Sapphire Rapids) (96 cores) 8 Intel Data Center GPU Max 1100 (Ponte Vecchio) 488 GB 418
pvc wait time: average 0.1 hours, range 0 to 6.9 hours over 30 days
pvc wall time: average 2.8 hours, range 0 to 24 hours over 30 days, wall-time limit 48h
bittware FPGA-based workloads and hardware acceleration (2 FPGA devices) 2 488 GB
nextsilicon Experimental NextSilicon accelerator workloads (restricted access. Utilizes a NextSilicon coprocessor. 2 Intel Xeon 8468 (Sapphire Rapids) (96 cores) 488 GB
nec Vector-engine architecture for vectorized HPC and MPI-based scientific computing. Utilizes a NEC Vector Engine, Type 20B-P card. 1 Intel Xeon 8268 (Cascade Lake) (48 cores) 760 GB
gh01 Grace-Hopper node for High-bandwidth AI/ML and HPC workloads requiring fast CPU-GPU data movement and memory interconnect 1 ARM Neoverse V2 (NVIDIA Grace ARM) (72 cores) 1 NVIDIA H100 488 GB

Software

No software usage data is currently reported for ACES in XDMoD.


Datasets

Name Description
pytorch-computer-vision-datasets

A collection of standard computer vision datasets formatted for PyTorch, supporting tasks like image classification and object detection. On ACES, these are used to benchmark GPU performance and test distributed deep learning workflows across accelerators.

pytorch-language-modelling-datasets

Text-based datasets for training NLP and language models in PyTorch. In ACES, they support benchmarking of large-scale, memory-intensive workloads and evaluating performance of transformer-based models across hardware.

tensorflow-computer-vision-datasets

Computer vision datasets optimized for TensorFlow, covering tasks such as classification and segmentation. Within ACES, they enable framework comparisons and validation of TensorFlow pipelines on heterogeneous accelerators.

tensorflow-language-modelling-datasets

NLP datasets prepared for TensorFlow, used for language modeling, translation, and text analysis. On ACES, they help evaluate distributed training performance and accelerator efficiency for sequential data workloads.

videollama_dataset

A multimodal dataset combining video and text for tasks like video understanding and captioning. In ACES, it is used to test high-throughput, multi-accelerator workflows and benchmark complex AI pipelines.


Storage

File System

Directory Path Quota Purge Backup Notes
$HOME /home/username 6 months after account deactivation Daily Small scripts, config files, not for general use
$SCRATCH /scratch/user/username 6 months after account deactivation or when quotas are exceeded. None Primary working directory for jobs, not for long-term storage.
$PROJECT /scratch/group/projectid 90 days after allocation expiration None Shared storage for group members

External Storage

Extra storage is available through Texas A&M HPRC: Google Drive (25GB free, with a paid expansion available), Microsoft OneDrive (25GB free), and HPRC Long Term Storage (paid dedicated storage for longer-term needs). For full details and current rates, see Texas A&M HPRC's Extra Storage Options guide: https://hprc.tamu.edu/kb/Helpful-Pages/Storage/

For data storage policies, please visit https://hprc.tamu.edu/kb/User-Guides/ACES/Policies/#data-storage


File Transfer

ACES supports several file transfer methods depending on your needs — Globus Connect is recommended for most transfers, but alternatives are available for more specific use cases. For details and setup instructions, see Texas A&M HPRC's File Transfer guide: https://hprc.tamu.edu/kb/Helpful-Pages/File-Transfer/

Note: FTP is not recommended, as it does not encrypt usernames, passwords, or data during transfer. Use SFTP whenever possible.

Supported Methods Data Transfer Node URL
GLOBUS | RECOMMENDED ACCESS TAMU ACES DTN https://app.globus.org/dashboard
SCP/SFTP ACCESS TAMU ACES DTN
FTP ACCESS TAMU ACES DTN
RSYNC ACCESS TAMU ACES DTN
RCLONE ACCESS TAMU ACES DTN
GDOWN ACCESS TAMU ACES DTN
PORTAL ACCESS TAMU ACES DTN https://portal.hprc.tamu.edu