Skip to content

Commit 5017750

Browse files
mmshadmmshad
authored
Merge pull request #352 from KempnerInstitute/351-fix-add-kempner_h200-details
adding h200 descriptions
2 parents 898b418 + 30126fb commit 5017750

4 files changed

Lines changed: 25 additions & 17 deletions

File tree

‎kempner_computing_handbook/s1_high_performance_computing/general_hpc_concepts/advanced_slurm_features.md‎

Lines changed: 4 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -46,7 +46,7 @@ salloc -p kempner_requeue --account=kempner_grads -N 1 -n 1 --gres=gpu:1 --cons
4646
You can find a list of features in Kempner AI cluster by,
4747

4848
```bash
49-
echo "PARTITION FEATURES" && sinfo -p kempner,kempner_requeue,kempner_h100 --noheader --format="%P %f" | sort | uniq
49+
echo "PARTITION FEATURES" && sinfo -p kempner,kempner_requeue,kempner_h100,kempner_h200 --noheader --format="%P %f" | sort | uniq
5050
```
5151

5252
and the result will be,
@@ -57,6 +57,7 @@ kempner intel,holyhdr,icelake,avx,avx2,avx512,gpu,a100,cc8.0
5757
kempner_requeue amd,holyndr,genoa,avx,avx2,avx512,gpu,h100,cc9.0
5858
kempner_requeue intel,holyhdr,icelake,avx,avx2,avx512,gpu,a100,cc8.0
5959
kempner_h100 amd,holyndr,genoa,avx,avx2,avx512,gpu,h100,cc9.0
60+
kempner_h200 amd,holyndr,genoa,avx,avx2,avx512,gpu,h200,cc9.0
6061
```
6162

6263
The following table describes each of the features in the Kempner AI cluster:
@@ -72,6 +73,7 @@ The following table describes each of the features in the Kempner AI cluster:
7273
| `avx`, `avx2`, `avx512` | Advanced Vector Extensions (AVX) are Intel's and AMD's SIMD (single instruction, multiple data) extensions for parallelism in computing. `avx512` offers more features and wider vector registers than `avx` and `avx2`. |
7374
| `gpu` | Indicates the presence of GPU resources in the partition. |
7475
| `h100` | NVIDIA H100 GPU, part of NVIDIA's Hopper architecture, designed for deep learning and high-performance computing tasks. |
76+
| `h200` | NVIDIA H200 GPU, part of NVIDIA's Hopper architecture, designed for deep learning and high-performance computing tasks. |
7577
| `a100` | NVIDIA A100 GPU, part of NVIDIA's Ampere architecture, also aimed at deep learning and high-performance computing. |
7678
| `cc9.0`, `cc8.0` | CUDA compute capability, indicating the version of the CUDA API and features the GPU supports. `cc9.0` and `cc8.0` refer to specific versions with different levels of support for CUDA features. |
7779

@@ -138,4 +140,4 @@ holygpu8a17504 3 32 1480
138140
holygpu8a11102 2 16 1430
139141
```
140142

141-
Using the information in the output table, you can adjust your job’s resource request (number of nodes, number of cores per node, and memory) to secure an allocation more quickly.
143+
Using the information in the output table, you can adjust your job’s resource request (number of nodes, number of cores per node, and memory) to secure an allocation more quickly.

‎kempner_computing_handbook/s1_high_performance_computing/general_hpc_concepts/understanding_slurm.md‎

Lines changed: 6 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -48,14 +48,16 @@ The following partitions are available for Kempner affiliates on the FASRC clust
4848

4949
| Partition | Description |
5050
|-----------|-------------|
51-
| `kempner` | This GPU block includes 2304 Intel Ice Lake cores and 144 Nvidia A100 40GB GPUs, with each water-cooled node featuring 64 cores, 1TB RAM, and 4 A100 GPUs, linked via HDR Infiniband, with a 7-day limit.|
52-
| `kempner_h100` | This GPU block includes 2304 AMD Genoa cores, 96 Nvidia H100 80GB GPUs, water-cooled nodes with 96 cores, 1.5TB RAM, and 4 H100 GPUs, interconnected via NDR Infiniband, with a 3-day limit. |
53-
| `kempner_requeue` | This partition utilizes `kempner` and `kempner_h100` partitions, designed for tasks that can be interrupted and restarted. This partition has a 7 day time limit.|
51+
| `kempner` | This GPU block includes 2304 Intel Ice Lake cores and 144 Nvidia A100 40GB GPUs, with each water-cooled node featuring 64 cores, 1TB RAM, and 4 A100 GPUs, linked via HDR Infiniband, with a 2-day limit.|
52+
| `kempner_h100` | This GPU block includes 2304 AMD Genoa cores, Nvidia H100 80GB GPUs, water-cooled nodes with 96 cores, 1.5TB RAM, and 4 H100 GPUs, interconnected via NDR Infiniband, with a 2-day limit. |
53+
| `kempner_h200` | This GPU block includes 2304 AMD Genoa cores, Nvidia H200 141GB GPUs, water-cooled nodes with 64 cores, 1.5TB RAM, and 4 H200 GPUs, interconnected via NDR Infiniband, with a 2-day limit. |
54+
| `kempner_rtx` | This GPU block includes 3072 AMD Turin cores, Nvidia RTX6000 96GB GPUs, air-cooled nodes with 128 cores, 1.5TB RAM, and 8 RTX6000 GPUs, interconnected via NDR Infiniband, with a 2-day limit. |
55+
| `kempner_requeue` | This partition utilizes `kempner`, `kempner_h100`, `kempner_h200`, and `kempner_rtx` partitions, designed for tasks that can be interrupted and restarted. This partition has a 7 day time limit.|
5456
| `kempner_dev` | This partition is dedicated to the engineering team and is not available to all users. |
5557
| `kempner_interactive` | This partition is dedicated to interactive jobs and lightweight development work. It provides 20GB A100 GPU slices created from full A100 40GB GPUs using NVIDIA Multi-Instance GPU (MIG) technology. |
5658

5759
```{note}
58-
The `kempner_interactive` partition is intended for interactive sessions, prototyping, debugging, and other lightweight development tasks, not for full-scale training or production runs. Because the GPUs are partitioned with MIG, each job receives a 20GB slice of an A100 rather than a full GPU. For resource-intensive jobs, use the `kempner` or `kempner_h100` partitions instead.
60+
The `kempner_interactive` partition is intended for interactive sessions, prototyping, debugging, and other lightweight development tasks, not for full-scale training or production runs. Because the GPUs are partitioned with MIG, each job receives a 20GB slice of an A100 rather than a full GPU. For resource-intensive jobs, use the `kempner`, `kempner_h100`, `kempner_h200`, or `kempner_rtx` partitions instead.
5961
```
6062

6163
````{tip}

‎kempner_computing_handbook/s1_high_performance_computing/kempner_cluster/kempner_policies_for_responsible_use.md‎

Lines changed: 8 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -12,16 +12,20 @@ We don’t limit jobs on the basis of size and encourage users to use capacity w
1212

1313
Please use the [Kempner requeue](https://docs.rc.fas.harvard.edu/kb/kempner-partitions/) partition if you want to try out something resource intensive but not urgent. This work may be requeued by higher priority work, so you should implement checkpointing in the event that something higher priority interrupts your run.
1414

15-
The cluster should never be used for CPU-only jobs. There are two types of Kempner Institute nodes on the cluster: Nodes with A100 40GB GPUs (`kempner` partition) and nodes with H100 80GB GPUs (`kempner_h100` partition). All nodes have 4 GPUs. Jobs should be submitted with **no more than**:
15+
The cluster should never be used for CPU-only jobs. There are four types of Kempner Institute nodes on the cluster: Nodes with A100 40GB GPUs (`kempner` partition), nodes with H100 80GB GPUs (`kempner_h100` partition), nodes with H200 141GB GPUs (`kempner_h200` partition), and nodes with RTX6000 96GB GPUs (`kempner_rtx`). The RTX nodes have 8 GPUs and other nodes have 4 GPUs. Jobs should be submitted with **no more than**:
1616

1717
- 16 cores and 240 GB per GPU for `kempner` partition, and
1818
- 24 cores and 360 GB per GPU for `kempner_h100` partition.
19+
- 16 cores and 360 GB per GPU for `kempner_h200` partition, and
20+
- 16 cores and 190 GB per GPU for `kempner_rtx` partition.
1921

2022
::::{important}
21-
Users should not submit jobs that utilize more than 10% of the cluster resources or that run for extended periods (exceeding 4 hours). Given the current resources, 10% of the cluster comprises:
23+
Users should not submit jobs that utilize more than the following cluster resources or that run for extended periods. Given the current resources, the limits in each partition are:
2224

23-
- 12 A100 GPUs (3 nodes) in the `kempner` partition, and
24-
- 12 H100 GPUs (3 nodes) in the `kempner_h100` partition.
25+
- 8 A100 GPUs across up to 4 nodes in the `kempner` partition, and a 2-day runtime limit.
26+
- 8 H100 GPUs across up to 4 nodes in the `kempner_h100` partition, and a 2-day runtime limit.
27+
- 8 H200 GPUs across up to 4 nodes in the `kempner_h200` partition, and a 2-day runtime limit.
28+
- 8 RTX6000 GPUs across up to 4 nodes in the `kempner_rtx` partition, and a 2-day runtime limit.
2529

2630
See the following guidelines for jobs that exceed this limit.
2731
::::

‎kempner_computing_handbook/s1_high_performance_computing/kempner_cluster/overview_of_kempner_cluster.md‎

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -15,13 +15,13 @@ In general terms, the cluster is reserved for individuals who are either associa
1515

1616
The following table provides an overview of the Kempner Institute AI cluster:
1717

18-
| Specification | A100 40GB | H100 80GB |
19-
|---------------------------|------------------------------------------------|-------------------------------------------------------|
20-
| **Total GPUs** | 144 | 384 |
21-
| **Servers (per rack)** | 36 | 24 |
22-
| **GPUs per Server** | 4 | 4 |
23-
| **CPU Cores per Server** | 64 | 96 |
24-
| **RAM per Server** | 1 TB | 1.5 TB |
18+
| Specification | A100 40GB | H100 80GB | H200 141GB | RTX6000 96GB |
19+
|---------------------------|------------------------------------------------|-------------------------------------------------------|-------------------------------------------------------|-------------------------------------------------------|
20+
| **Total GPUs** | 144 | 384 | 384 | 192 |
21+
| **Servers (per rack)** | 36 | 24 | 24 | 6 |
22+
| **GPUs per Server** | 4 | 4 | 4 | 8 |
23+
| **CPU Cores per Server** | 64 | 96 | 64 | 128 |
24+
| **RAM per Server** | 1 TB | 1.5 TB | 1.5 TB | 1.5 TB |
2525

2626

2727
```{figure} figures/svg/cluster_diagram.svg

0 commit comments

Comments
 (0)