Hardware Information¶

The following hardware summaries may be useful for selecting partitions for workflows and for grant proposal writing. If any information is missing that would be helpful to you, please be sure to contact us or create an issue on our tracker.

Tip

The tables in this section are wide and can be scrolled horizontally to display more information.

Cheaha HPC Cluster¶

Summary¶

The table below contains a summary of the computational resources available on Cheaha and relevant Quality of Service (QoS) Limits. QoS limits allow us to balance usage and ensure fairness for all researchers using the cluster. QoS limits are not a guarantee of resource availability.

In the table, Slurm partitions are grouped by shared QoS limits on cores, memory, and GPUs. Node limits are applied to partitions independently. All limits are applied to researchers independently.

Examples of how to make use of the table:

Suppose you submit 30 jobs to the "express" partition, and suppose each job needs 10 cores each. Hypothetically, in order for all of the jobs to start at once, 300 cores would be required. The QoS limit on cores is 264 on the "express" partition, so at most 26 jobs (260 cores) can start at once. The remaining 4 jobs will be held in queue, because starting one more would go beyond the QoS limit (270 > 264).
Suppose you submit 5 jobs to the "medium" partition and 5 to the "long" partition, each requiring 1 node. Then, 10 total nodes would be needed. In this case, it is possible that all 10 jobs can start at once because partition node limits are separate. If all 5 jobs start, jobs on the "medium" partition.
Suppose you submit 5 jobs to the "amperenodes" partition and 5 to "amperenodes-medium", for a total of 10 A100 GPUs. Additionally, you also submit 4 jobs to the "pascalnodes" partition totaling 8 P100 GPUs. Then 4 of the "gpu: ampere" group jobs can start at once, because the QoS limit is 4 GPUs there. Additionally, all 4 of the "gpu: pascal" group jobs, because the QoS limit is 8 GPUs there. In this case, the QoS for each group is separate.

Partition	Time Limit in Hours	Nodes (Limit/Partition)	Cores/Node (Limit/Person)	Mem GB/Node (Limit/Person)	GPU/Node (Limit/Person)
cpu: amd
amd-hdr100	150	34 (5)	128 (264)	504 (3072)

cpu: intel
express	2	51 (~)	48 (264)	754 (3072)
short	12	51 (44)	48 (264)	754 (3072)
medium	50	51 (44)	48 (264)	754 (3072)
long	150	51 (5)	48 (264)	754 (3072)

gpu: ampere
amperenodes	12	20 (TBD)	32 (64)	189 (384)	2 (4)
amperenodes-medium	48	20 (TBD)	32 (64)	189 (384)	2 (4)

gpu: pascal
pascalnodes	12	18 (~)	28 (56)	252 (500)	4 (8)
pascalnodes-medium	48	7 (~)	28 (56)	252 (500)	4 (8)

mem: large
largemem	50	13 (10)	24 (290)	755 (7168)
largemem-long	150	5 (10)	24 (290)	755 (7168)

The full table can be downloaded here.

Information about GPU efficiency can be found at Making the Most of GPUs.

Details¶

Detailed hardware information, including processor and GPU makes and models, core clock frequencies, and other information for current hardware are in the table below.

Generation	Compute Type	Total Cores	Total Memory Gb	Total Gpus	Cores Per Node	Cores Per Die	Dies Per Node	Die Brand	Die Name	Die Frequency Ghz	Memory Per Node Gb	Gpu Per Node	Gpu Brand	Gpu Name	Gpu Memory Gb	Nodes
1	cpu: amd	128	1024		2	1	2	AMD	Opteron 242	1.6	16					64
2	cpu: intel	192	1152		8	4	2	Intel	Xeon E5450	3	48					24
3	cpu: intel	384	1536		12	6	2	Intel	Xeon X5650	2.66	48					32
3	cpu: intel	192	1536		12	6	2	Intel	Xeon X5650	2.66	96					16
4	cpu: intel	48	1152		16	8	2	Intel	Xeon X5650	2.7	384					3
5	cpu: intel	192	1152		16	8	2	Intel	Xeon E2650	2	96					12
6	cpu: intel	336	5376		24	12	2	Intel	Xeon E5-2680 v3	2.5	384					14
6	cpu: intel	912	9728		24	12	2	Intel	Xeon E5-2680 v3	2.5	256					38
6	cpu: intel	1056	5632		24	12	2	Intel	Xeon E5-2680 v3	2.5	128					44
7	gpu: pascal	504	4608	72	28	14	2	Intel	Xeon E5-2680 v4	2.4	256	4	NVIDIA	Tesla P100	16	18
8	cpu: intel	504	4032		24	12	2	Intel	Xeon E5-2680 v4	2.5	192					21
8	mem: large	240	7680		24	12	2	Intel	Xeon E5-2680 v4	2.5	768					10
8	mem: large	96	6144		24	12	2	Intel	Xeon E5-2680 v4	2.5	1536					4
9	cpu: intel	2496	39936		48	24	2	Intel	Xeon Gold 6248R	3	768					52
10	cpu: amd	4352	17408		128	64	2	AMD	Epyc 7713 Milan	2	512					34
11	gpu: ampere	2560	10240	40	128	64	2	AMD	Epyc 7763 Milan	2.45	512	2	NVIDIA	A100	80	20
1	cpu: intel	240	960		48	12	4	Intel	Xeon Gold 6248R	3	192					5
1	gpu: ampere	512	4096	32	128	64	2	AMD	Epyc 7742 Rome	2.25	1024	8	NVIDIA	A100	40	4
1	cpu: intel	144	576		48	12	4	Intel	Xeon Gold 6248R	3	192					3
1	gpu: ampere	512	4096	32	128	64	2	AMD	Epyc 7742 Rome	2.25	1024	8	NVIDIA	A100	40	4

The full table can be downloaded here.

The table below is a theoretical analysis of FLOPS (floating point operations per second) based on processor instructions and core counts, and is not a reflection of efficiency in practice.

Generation	Cpu Tflops Per Node	Gpu Tflops Per Node	Tflops Per Node	Nodes	Tflops
7	1.08	17.06	18.14	18	326.43
8	0.96		0.96	21	20.16
8	0.96		0.96	10	9.60
8	0.96		0.96	4	3.84
9	2.30		2.30	52	119.81
10	4.10		4.10	34	139.26
11	5.02	15.14	20.15	20	403.10
Total					1,022.20

The full table can be downloaded here.

For information on using Cheaha, see our dedicated section.

Cloud Service at cloud.rc¶

The Cloud service hardware consists of 5 Intel nodes and 4 DGX-A100 nodes. A description of the available hardware are summarized in the following table.

Fabric	Generation	Compute Type	Total Cores	Total Memory Gb	Total Gpus	Cores Per Node	Memory Per Node Gb	Nodes	Cpu Info	Gpu Info
cloud	1	cpu	240	960		48	192	5	Intel Xeon Gold 6248R 3.00 GHz
cloud	1	gpu	512	4096	32	128	1024	4	AMD Epyc 7742 Rome 2.25 GHz	NVIDIA A100 40 GB
Total			752	5056	32			9

The full table can be downloaded here.

The table below is a theoretical analysis of FLOPS (floating point operations per second) based on processor instructions and core counts, and is not a reflection of efficiency in practice.

Generation	Cpu Tflops Per Node	Gpu Tflops Per Node	Tflops Per Node	Nodes	Tflops
1	2.30		2.30	5	11.52
1	4.61	77.97	82.58	4	330.3
Total					341.82

The full table can be downloaded here.

For information on using our Cloud service at cloud.rc, see our dedicated section.

Kubernetes Container Service¶

Important

The Kubernetes fabric is still in deployment and not ready for researcher use. We will be sure to inform you when the service is ready. The following information is planned hardware.

The Kubernetes container service hardware consists of 5 Intel nodes and 4 DGX-A100 nodes. A description of the available hardware are summarized in the following table.

Fabric	Generation	Compute Type	Total Cores	Total Memory Gb	Total Gpus	Cores Per Node	Memory Per Node Gb	Nodes	Cpu Info	Gpu Info
container	1	cpu	144	576		48	192	3	Intel Xeon Gold 6248R 3.00 GHz
container	1	gpu	512	4096	32	128	1024	4	AMD Epyc 7742 Rome 2.25 GHz	NVIDIA A100 40 GB
Total			656	4672	32			7

The full table can be downloaded here.

The table below is a theoretical analysis of FLOPS (floating point operations per second) based on processor instructions and core counts, and is not a reflection of efficiency in practice.

Generation	Cpu Tflops Per Node	Gpu Tflops Per Node	Tflops Per Node	Nodes	Tflops
1	2.30		2.30	3	6.91
1	4.61	77.97	82.58	4	330.3
Total					337.21

The full table can be downloaded here.