Stampede3

Description

Stampede3 is generously funded through the National Science Foundation and is designed to serve today's researchers as well as support the research community on an evolutionary path toward many-core processors and accelerated technologies. Stampede 3 maintains the familiar programming model for all of today's users, and thus will be broadly useful for traditional simulation users, users performing data intensive computations, and emerging classes of new users.

Resource ID
1923
Global Resource ID
stampede3.tacc.access-ci.org
Resource Type
Compute
Latest Status
production
Latest Status Begin
Latest Status End
Project Affiliation
ACCESS
Organization Name
Texas Advanced Computing Center
RP Description

Stampede3 is a compute cluster consisting of Intel Sapphire Rapids high-bandwidth-memory nodes, Ice Lake CPU nodes, large-memory Ice Lake NVDIMM nodes, NVIDIA H100 GPU nodes, Intel Data Center GPU Max nodes, and Skylake CPU nodes. It is particularly well suited for large-scale simulation, high-throughput computing, artificial intelligence, GPU-accelerated workloads, and large-memory applications, and is often used for groundbreaking open science research projects.

MFA Required
On
External Storage

Every Stampede3 allocation comes with 2 TB on Ranch, TACC's long-term archival tape system, at no additional cost, and it does not consume ACCESS credits. Ranch is for final data that is unlikely to change and does not need frequent access. Move data to it with Globus, or with scp myfile ${ARCHIVER}:${ARCHIVE}/myfilepath. See [Ranch User Guide].

Projects that need more space than $HOME and $WORK provide can apply for Corral, TACC's online project storage system, which Stampede3 reaches through the same transfer methods. See [Corral User Guide].

File Transfer Text

Use [Globus Data Transfer Guide at TACC] for datasets larger than about 200 GB, and the SSH tools scp, rsync and sftp for anything smaller - see [SSH-based Data Transfer].

TACC does not publish a separate data transfer node for Stampede3: command-line transfers go to stampede3.tacc.utexas.edu, the same host you log in to. Keep those transfers small, because the login nodes are shared, and move large data with Globus instead.

File Transfer Methods
Transfer Method
Globus
Data Transfer Node / Globus Collection
TACC Stampede3
Transfer Method
scp
Data Transfer Node / Globus Collection
stampede3.tacc.utexas.edu
Transfer Method
rsync
Data Transfer Node / Globus Collection
stampede3.tacc.utexas.edu
Transfer Method
sftp
Data Transfer Node / Globus Collection
stampede3.tacc.utexas.edu
Storage Text

Stampede3 has three file systems: a small, backed-up $HOME for source code and configuration, $WORK on TACC's shared Stockyard file system for data you want to keep, and a large $SCRATCH for job output. Run your jobs out of $SCRATCH rather than $HOME.

Two things surprise people. The 1 TB $WORK quota is shared across every TACC system you use, not granted separately on each one, and files on $SCRATCH are purged once they have not been accessed for ten days. Only $HOME is backed up. See [Stampede3 Managing Your Files] for the details.

Jobs Information

You can run jobs at different sizes and durations on Stampede3. The following lists the different queues that you can submit to, describing how many nodes you get, how long you can run, the type of resources you get, and the average wait time.

Stampede3 uses the Slurm scheduler and supports multi-node and MPI jobs. Nothing is assumed for you: your job script must set --partition, --time, --account and --ntasks explicitly. See [Stampede3 Running Jobs] for submission commands, sample job scripts and job management.

TACC does not share nodes between users on any of its compute resources, so the smallest thing you can request is a whole node and a job that uses one core is charged for all of it. There is no shared or interactive partition; skx-dev is the development queue and is capped at 16 nodes and 2 hours, while every other queue allows up to 48 hours.

Stampede3 is five different node types behind one scheduler, so choosing a queue is choosing the hardware. The spr nodes carry 128 GB of on-package high-bandwidth memory and no DDR, so work that needs more memory per core belongs on icx or skx; nvdimm is single-node only and exists for jobs that fit nowhere else; and the two GPU queues take at most 4 nodes per job, with pvc providing Intel Data Center GPU Max GPUs rather than NVIDIA ones, so CUDA binaries will not run there. Per-queue limits are listed in [Stampede3 Production Queues].

Jobs are charged in node-hours on the nodes reserved, not on the cores used, and the rate differs by queue. Run /usr/local/etc/taccinfo on a login node to see your project balances and disk quotas, and see [Stampede3 Job Accounting] for the billing formula and current rates.

Storage Filesystems
Directory
Home
File System Path
$HOME
Quota size
15
Quota inode amount
300000
Purge Policy
After account deactivation
Backup Policy
Backed up regularly
Notes
VAST filesystem. Keep source code, scripts and configuration here; run jobs from $SCRATCH.
Directory
Work
File System Path
$WORK
Quota size
1000
Quota inode amount
3000000
Purge Policy
Not on a schedule
Backup Policy
Not backed up
Notes
Lustre filesystem on TACC's Stockyard. The 1 TB quota is shared across all TACC systems, not granted per system. Not intended for parallel or high-intensity file operations.
Directory
Scratch
File System Path
$SCRATCH
Purge Policy
Subject to purge if access time is more than 10 days old.
Backup Policy
Not backed up
Notes
VAST filesystem. Main working area - write job output here rather than to $HOME
Queue Specifications
Queue Name
nvdimm
Purpose
Jobs that need more memory than any other Stampede3 node can offer; one node per job
CPU Type
2x Intel Xeon Platinum 8380
GPU Type
None
GPU Count
0
GPU vRAM
0
CPU Count
80
Node RAM
4000
Queue Name
pvc
Purpose
Workloads built for Intel GPUs; up to 4 nodes per job
CPU Type
2x Intel Xeon Platinum 8480
GPU Type
Intel Data Center GPU Max 1550
GPU Count
4
GPU vRAM
124
CPU Count
96
Node RAM
1000
Queue Name
h100
Purpose
CUDA and deep learning workloads on NVIDIA GPUs; up to 4 nodes per job
CPU Type
2x Intel Xeon Platinum 8468
GPU Type
NVIDIA H100 SXM5
GPU Count
4
GPU vRAM
96
CPU Count
96
Node RAM
4000
Queue Name
icx
Purpose
General-purpose CPU jobs that need more memory per core than the SPR nodes provide; up to 32 nodes per job
CPU Type
2x Intel Xeon Platinum 8380
GPU Type
None
GPU Count
0
GPU vRAM
0
CPU Count
80
Node RAM
256
Queue Name
spr
Purpose
Memory-bandwidth-bound jobs that benefit from on-package high-bandwidth memory; up to 32 nodes per job
CPU Type
2x Intel Xeon CPU MAX 9480
GPU Type
None
GPU Count
0
GPU vRAM
0
CPU Count
112
Node RAM
128
Queue Name
skx
Purpose
General-purpose CPU jobs, including the largest multi-node runs; up to 256 nodes per job
CPU Type
2x Intel Xeon Platinum 8160
GPU Type
None
GPU Count
0
GPU vRAM
0
CPU Count
48
Node RAM
192
Queue Name
skx-dev
Purpose
Development and testing on Skylake nodes; short jobs only, up to 16 nodes and 2 hours
CPU Type
2x Intel Xeon Platinum 8160
GPU Type
None
GPU Count
0
GPU vRAM
0
CPU Count
48
Node RAM
192
Datasets