GPU Nodes

The IBEX cluster contains different architectures of GPUs like Turing, Pascal and Volta. These GPUs are further described below for your source code compilation and job submission.

The IBEX cluster has 62 GPU compute nodes (392 GPU cards)  and it’s summarized in Table 1. These various GPUs are accessed by the SLURM scheduling using the constraints "--gres=gpu:GPUTYPE:x”, where x is for number of GPUs.

For example, “--gres=gpu:gtx1080ti:4” allocates 4 GTX GPUs.

 

                                        Table 1. List of GPU architectures in IBEX Cluster

Sl. No

GPU Architecture

Available GPU Cards Per Node

Available Number of Nodes

GPU Memory (Per Card)

Usable Node Memory*

 CPU Type CPU Core Count (Per Node)

Constraint for SLURM Scheduling

1.

Turing:

RTX 2080 TI

8

4

12GB

366GB

 
Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz
 32

"--gres=gpu:rtx2080ti:1"

2.

Pascal:

GTX 1080 TI

4 or 8

12
(8*4 and 4*8)

12GB

246GB
or
366GB

 
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz

(add --constraint cpu_intel_e5_2699_v3)

or

Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz

(add --constraint cpu_intel_gold_6142)

36

or

32

"--gres=gpu:gtx1080ti:1"

3.

Pascal:

P100

4

5

16GB

246GB

 
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz
 36

"--gres=gpu:p100:1"

4.

Pascal:

P6000

2

2

22GB

246GB

 
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz
 36

"--gres=gpu:p6000:1"

5.

Volta:

V100

4 or 8

38

(8*4 and 30*8)

32GB

366GB

or

745GB

 
Intel(R) Xeon(R) Gold 6142 CPU @ 2.60GHz

(add --constraint cpu_intel_gold_6142)

or

Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz

(add --constraint cpu_intel_platinum_8260)

32

or

48

"--gres=gpu:v100:1"

Note1: The allocation of CPU memory can be done with `--mem=###G` constraint in SLURM job scripts. The amount of memory depends on the job characteristization. A good starting place would be at least as much as the GPU memory they will use. For example: 2 x v100 GPUs would allocate at least `--mem=64G` for the CPUs.

Note2: The glogin node has a single Quadro K6000 GPU for compilation of the source code.

* The usable node memory represents the available memory for job execution.

 

More on Slurm Constraints:

"ref_32T" and "gpu_ai" are used to differentiate the newer generation of the V100 GPU nodes from the old ones.
The new nodes have 32TB of NVMes as local storage. And some ML reference DBs have been copied to those NVMes to enhance jobs performance instead of using the shared BeeGFS scratch.

 

Slurm Partition:

Continuous efforts has been made for fair share allocation of resources on Ibex, the following partitions has been implemented seamlessly to our users.

gpu_wide for jobs with 4+ gpus per node

gpu_wide24 wide jobs with time limit less than 24 hours

gpu4  for short GPU jobs (less than 4 hours)

Note: Users can't specify those partitions in their scripts. This is done automatically by SLURM.
 
 
 
 
For further info or queries send us using the Contact Us page.

Alternatively, send an email to ibex@hpc.kaust.edu.sa.