COSMA GPU Nodes

Last updated: 21/01/2026

Summary of Available Nodes

GPU	Count	Node(s)	Access	Partition	Account	Interconnect	Status
V100	6x	gn001	SSH	-	do016	-	✓
A30	8x total	gc001-008	Slurm	dine2	do015	CerIO (composable)	Draining
A100	3x total	mad04-06	Slurm / SSH	cosma8-shm	do016	Liqid (composable)	✓
GH200	2x	gn002-003	Slurm / SSH	gracehopper	do016	-	✓
H100 NVL	1x	gn004	SSH	intelhopper (down)	do016	-	✓ (SSH only)
MI100	1x	ga004	Slurm	cosma8-shm2	do018	-	✓
MI210	2x each	ga005-006	Slurm	cosma8-shm2	do018	-	ga006 draining
MI300X	8x	ga007	Slurm	mi300x	do018	-	✓
MI300A	4x	ga008	SSH	mi300a (down)	do018	-	Down
PVC	2x	gi001	-	-	do017	-	Dead

GPU Specifications (Theoretical)

GPU	Architecture	FP64 TFLOPS	VRAM (GB)	Memory BW (GB/s)	Backend
V100	Volta	7.0	32	900	CUDA
A30	Ampere	5.2	24	933	CUDA
A100	Ampere	9.7	40	1555	CUDA
GH200	Hopper	34	96	4000	CUDA
H100 PCIe	Hopper	26	80	2000	CUDA
H100 SXM	Hopper	34	80	3350	CUDA
H100 NVL	Hopper	34	94	3938	CUDA
MI100	CDNA	11.5	32	1200	HIP
MI210	CDNA2	22.6	64	1600	HIP
MI300X	CDNA3	81.7	192	5300	HIP
MI300A	CDNA3 (APU)	61.3	128	5300	HIP
PVC	Xe HPC	52.4	128	3280	SYCL

Detailed Node Information

NVIDIA CUDA Systems

Node	GPUs	Host	Access	Account	Arch	Notes
gn001	10x V100 (32GB)	768GB, 2x Xeon Gold 5218	SSH	do016?	sm_70
gc001-008	0-8x A30 (composable)	2TB, 64 cores (Sapphire Rapids)	dine2	do015	sm_80	CerIO fabric; 8 GPUs total across 8 nodes
mad04, mad05	0-3x A100 40GB (composable)	4TB, 128 cores	cosma8-shm	do016?	sm_80	Liqid fabric; default 1 GPU each
mad06	0-3x A100 (composable)	1TB, 128 cores (Milan-X, 768MB L3)	SSH	do016?	sm_80
gn002	GH200	ARM (aarch64)	SSH	do016	sm_90	ARM host: requires native compilation
gn003	GH200	ARM (aarch64)	gracehopper	do016	sm_90	No system cmake; CUDA 13.0 at /usr/local/cuda-13.0/
gn004	1x H100 PCIe	x86 (Intel)	SSH	do016?	sm_90

AMD HIP/ROCm Systems

Node	GPUs	Host	Access	Account	Arch	Notes
ga004	1x MI100	1TB, 128 cores (Milan 7713)	cosma8-shm2	do018?	gfx908
ga005, ga006	2x MI210 each	1TB, 64 cores (EPYC 7513)	cosma8-shm2	do018?	gfx90a	Use --exclude=ga004 for MI210
ga007	8x MI300X	-	mi300x	do018	gfx942
ga008	4x MI300A	500GB shared	SSH	do018	gfx942	APU: CPU/GPU share physical RAM

Note: ROCm 6.3.0 installed system level on AMD nodes at /opt/rocm-6.3.0/.

Intel Systems

Node	GPUs	Access	Account	Status
gi001	2x Ponte Vecchio	-	do017	Dead

Note: Likely to stay dead for now!

Compilers

System	Compiler	Location	Notes
CUDA (x86, sm_80+)	nvcc 13.0	module load nvhpc/25.11	sm_80, sm_90
CUDA (GH200)	nvcc 13.0	/usr/local/cuda-13.0/bin/nvcc	ARM host: pip3 install --user cmake
HIP (AMD nodes)	hipcc 6.3.0	/opt/rocm-6.3.0/bin/hipcc	System-level: no module needed

Note: The hipcc/6.0amd, hipcc/6.3amd, hipcc/7.0amd modules only set CC/CXX environment variables for dependent modules (e.g. MPI). ROCm is already available at system level.

Project Codes (from SAFE)

Code	Systems
do015	DINE2 (A30s)
do016	NVIDIA GPUs (V100, A100, GH200, H100), cosma8-shm
do017	Intel GPUs (dead)
do018	AMD GPUs (MI100, MI210, MI300X, MI300A)
do009	DINE, miscellaneous

Access/Useful Commands

Check partition and node status

sinfo -a # list all partitions and statuses
sinfo -p cosma8-shm2 -N -l # detailed node info for a partition
squeue -p mi300x # show jobs running on partition
scontrol show partition=<partition_name> | grep AllowAccounts

Check access permissions

id # show groups you belong to
scontrol show partition= | grep AllowAccounts # show accounts that can use partitions

GPU Availability:

# NVDIA nodes
nvidia-smi

# AMD nodes
rocm-smi
rocminfo

Can use in slurm script to check GPU present.

Slurm submission

# by partition
sbatch -p cosma8-shm -A do016 script.sh # A100 on cosma8-shm
# by specific node
sbatch --nodelist=ga004 script.sh # ga004

# interactive session
srun -p mi300x -A do018 -t 10 --pty /bin/bash # mi300x

Direct SSH access

# From a COSMA login node:
ssh gn001 # V100s
ssh gn002 # Grace-Hopper (ARM)
ssh gn004 # H100
ssh ga008 # MI300A
ssh mad06 # A100 + Milan-X