From c04f8dafc7b209983b0363f81ce22d1d4d7cfcfe Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 10:38:58 -0600 Subject: [PATCH 01/13] Update Matlab page with GPU compatibility notes --- docs/software/matlab.md | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/docs/software/matlab.md b/docs/software/matlab.md index fcd267864..85e75ef25 100644 --- a/docs/software/matlab.md +++ b/docs/software/matlab.md @@ -5,6 +5,12 @@ resources. In this tutorial we will learn how to run Matlab on these resources. The tutorial assumes you are familiar with Matlab and basic Linux terminal commands. +```{attention} +If you are running on **A100, L40, or H200 GPUs**, use Matlab R2025b. + +If you are running on an **RTX Pro 6000 GPU**, use Matlab R2025b_FC to ensure forward compatibility with the GPU compute capability. Future versions of Matlab are anticipated to work on all GPUs without needing a forward compatibility module. +``` + There are two basic ways to run Matlab (or many other kinds of software) on RC resources. The first is through an interactive job, and the second is through a batch job. An interactive job allows one From cac83fd3686caddd242d29132bc845fbf4e710c1 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 10:52:08 -0600 Subject: [PATCH 02/13] Update AlphaFold documentation with GPU support details Added GPU compatibility notes for AlphaFold 2 and 3. --- docs/software/alphafold.md | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/docs/software/alphafold.md b/docs/software/alphafold.md index 05f9e29c2..dded738e9 100644 --- a/docs/software/alphafold.md +++ b/docs/software/alphafold.md @@ -10,6 +10,10 @@ AlphaFold is a program that predicts the three-dimensional structure of proteins ````{tab-item} AlphaFold 2 :sync: batch-scripting-ex1 +```{attention} +AlphaFold 2 is supported on A100 and L40 GPUs, but **not on H200 and RTX Pro 6000 GPUs**. +``` + Load the default AlphaFold 2 module: ``` module load alphafold/2.3.1 @@ -86,6 +90,11 @@ AlphaFold 3 has a substantially updated diffusion-based architecture that is cap On CURC’s Alpine system, AlphaFold 3 is available as a containerized module. It uses Apptainer/Singularity under the hood and is fully self-contained except for the separately downloaded model parameters (required). #### AlphaFold 3 Module +```{attention} +AlphaFold 3 versions 3.0.0 and 3.0.1 are supported on A100, L40, and H200 GPUs, but **not on RTX Pro 6000 GPUs**. +AlphaFold 3 version 3.0.3 is supported on **all NVIDIA GPU types** on Alpine. +``` + Load AlphaFold 3 module: ``` module load alphafold/3.0.0 @@ -186,4 +195,4 @@ export AF3_MODEL_PARAMETERS_DIR=/path/to/alphafold3/params run_alphafold --json_path=$INPUT_FILE --output_dir=$OUTPUT_DIR --model_dir=$AF3_MODEL_PARAMETERS_DIR ``` ```` -````` \ No newline at end of file +````` From 7236353fc492f2f8b899f33481333a4a592dec92 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 10:56:52 -0600 Subject: [PATCH 03/13] Add FAQ on NVIDIA SDK version for Alpine GPUs Added a new FAQ entry regarding the recommended NVIDIA SDK version for Alpine GPUs, including specific GPU architectures. --- docs/getting_started/faq.md | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/docs/getting_started/faq.md b/docs/getting_started/faq.md index 9ee8fa127..eb1dd1a0e 100644 --- a/docs/getting_started/faq.md +++ b/docs/getting_started/faq.md @@ -351,6 +351,18 @@ Before submitting a software installation request, please review our [Software i For individuals who need to install Python libraries not included in our base Python or Anaconda modules, we recommend using Conda environments through the Anaconda module. Instructions for creating a custom Conda environment can be found on our [Python and R with Anaconda](../software/python.md) documentation page. :::: +### Which NVIDIA SDK version should I use on Alpine GPUs? +::::{dropdown} Show +:icon: note + +Our recommendation is to use >= nvhpc_sdk/2025.251 on all GPUs. If you require more specific information on the CUDA versions a given GPU architecture will support, please see the [NVIDIA Toolkit, Driver and Architecture Matrix](https://docs.nvidia.com/datacenter/tesla/drivers/cuda-toolkit-driver-and-architecture-matrix.html), noting the following architectures available on Alpine: + +* A100: Ampere +* L40: Ada +* H200: Hopper +* RTX Pro 6000: Blackwell +:::: + ## Alpine ### Why do I get an `Invalid Partition` error when running an Alpine job? From c993f8a19e0f43b72c25ad6607a0f9c31a82d968 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 11:19:04 -0600 Subject: [PATCH 04/13] Add warning for CUDA version compatibility to NVIDIA profiling page Added a warning about CUDA version compatibility for NVIDIA GPUs. --- docs/programming/profiling-nvidia-gpu-performance.md | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/docs/programming/profiling-nvidia-gpu-performance.md b/docs/programming/profiling-nvidia-gpu-performance.md index 406c8f32b..af5407e2e 100644 --- a/docs/programming/profiling-nvidia-gpu-performance.md +++ b/docs/programming/profiling-nvidia-gpu-performance.md @@ -1,5 +1,9 @@ # Profiling NVIDIA GPU Performance +```{warning} +Ensure you are using a CUDA version compatible with the GPU you are using. For more information, see our FAQ on [Which NVIDIA SDK version should I use on Alpine GPUs?](../getting_started/faq.md#which-nvidia-sdk-version-should-i-use-on-alpine-gpus) +``` + The NVIDIA Performance Counters provide low-level metrics on GPU usage, enabling users to understand how efficiently their code uses the GPU. This is especially important for optimizing workloads on Alpine’s A100 GPU nodes, where GPU time is a valuable and shared resource. The following tools are available for interacting with performance counters: @@ -1154,4 +1158,4 @@ Time (%) Total Time (ns) Num Calls Name This output means that most of the runtime was spent in system calls like `poll` and `ioctl`, which are unrelated to the kernel execution. This suggests the application is not compute-bound, and system overhead is prominent due to the small workload. ```` -````` \ No newline at end of file +````` From aaa318f6b872508151d269ea58da0a843ec3975c Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 11:29:34 -0600 Subject: [PATCH 05/13] Remove A100 specific-MIG references on profiling page --- docs/programming/profiling-nvidia-gpu-performance.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/programming/profiling-nvidia-gpu-performance.md b/docs/programming/profiling-nvidia-gpu-performance.md index af5407e2e..241b3f377 100644 --- a/docs/programming/profiling-nvidia-gpu-performance.md +++ b/docs/programming/profiling-nvidia-gpu-performance.md @@ -245,7 +245,7 @@ Example output of `nvidia-smi` on the `aa100` partition using [matrixMultiply.cu ### nvidia-smi on MIG-Enabled GPUs -Some A100 GPUs on our systems are MIG-enabled (Multi-Instance GPU). On these nodes, `nvidia-smi` shows a different output format, displaying information for both full GPUs and individual MIG instances. +Some GPUs on our systems are MIG-enabled (Multi-Instance GPU). On these nodes, `nvidia-smi` shows a different output format, displaying information for both full GPUs and individual MIG instances. Here's an example output from a MIG-enabled A100 node: ``` @@ -411,7 +411,7 @@ $ ncu --set full --target-processes all ./vectorAdd - `--target-processes all`: Profiles all child processes (useful for multi-threaded applications). ```{note} -`ncu` is not compatible with MIG-enabled GPUs. Ensure you run `ncu` only on A100 nodes without MIG. +`ncu` is not compatible with MIG-enabled GPUs. Ensure you run `ncu` only on GPU nodes without MIG. ``` ::::{dropdown} Click here to view the full `ncu` report From 8a32e30ff65b3c6ca68e2979362a16e015f9dc0b Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 11:33:58 -0600 Subject: [PATCH 06/13] Minor formatting fix --- docs/getting_started/faq.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/getting_started/faq.md b/docs/getting_started/faq.md index eb1dd1a0e..3eb408d8c 100644 --- a/docs/getting_started/faq.md +++ b/docs/getting_started/faq.md @@ -355,7 +355,7 @@ For individuals who need to install Python libraries not included in our base Py ::::{dropdown} Show :icon: note -Our recommendation is to use >= nvhpc_sdk/2025.251 on all GPUs. If you require more specific information on the CUDA versions a given GPU architecture will support, please see the [NVIDIA Toolkit, Driver and Architecture Matrix](https://docs.nvidia.com/datacenter/tesla/drivers/cuda-toolkit-driver-and-architecture-matrix.html), noting the following architectures available on Alpine: +Our recommendation is to use >= `nvhpc_sdk/2025.251` on all GPUs. If you require more specific information on the CUDA versions a given GPU architecture will support, please see the [NVIDIA Toolkit, Driver and Architecture Matrix](https://docs.nvidia.com/datacenter/tesla/drivers/cuda-toolkit-driver-and-architecture-matrix.html), noting the following architectures available on Alpine: * A100: Ampere * L40: Ada From 4c59a3cd78490d67575f5a0d28daa619a33a4e62 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 12:53:34 -0600 Subject: [PATCH 07/13] Apply suggestions from code review Co-authored-by: monaghaa <33913248+monaghaa@users.noreply.github.com> --- docs/programming/profiling-nvidia-gpu-performance.md | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/docs/programming/profiling-nvidia-gpu-performance.md b/docs/programming/profiling-nvidia-gpu-performance.md index 241b3f377..885a77c27 100644 --- a/docs/programming/profiling-nvidia-gpu-performance.md +++ b/docs/programming/profiling-nvidia-gpu-performance.md @@ -4,7 +4,7 @@ Ensure you are using a CUDA version compatible with the GPU you are using. For more information, see our FAQ on [Which NVIDIA SDK version should I use on Alpine GPUs?](../getting_started/faq.md#which-nvidia-sdk-version-should-i-use-on-alpine-gpus) ``` -The NVIDIA Performance Counters provide low-level metrics on GPU usage, enabling users to understand how efficiently their code uses the GPU. This is especially important for optimizing workloads on Alpine’s A100 GPU nodes, where GPU time is a valuable and shared resource. +The NVIDIA Performance Counters provide low-level metrics on GPU usage, enabling users to understand how efficiently their code uses the GPU. Optimizing workloads on Alpine’s GPU nodes is important due to the high demand for these valuable shared resources. The following tools are available for interacting with performance counters: @@ -374,7 +374,7 @@ NVIDIA Nsight Compute is a command-line CUDA kernel profiler that provides detai ### Why do these metrics matter? -Modern GPUs, such as the NVIDIA A100, have hundreds of compute units (Streaming Multiprocessors or SMs). To fully exploit this parallel architecture, your kernel must be configured to launch enough threads and blocks to keep these units busy. +Modern GPUs have hundreds of compute units (Streaming Multiprocessors or SMs). To fully exploit this parallel architecture, your kernel must be configured to launch enough threads and blocks to keep these units busy. Key Features: @@ -410,7 +410,6 @@ $ ncu --set full --target-processes all ./vectorAdd - `--target-processes all`: Profiles all child processes (useful for multi-threaded applications). -```{note} `ncu` is not compatible with MIG-enabled GPUs. Ensure you run `ncu` only on GPU nodes without MIG. ``` From 1f1af79d118a027d5b1d934005bc67b7213ca29a Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 13:06:11 -0600 Subject: [PATCH 08/13] Move ncu mig note --- docs/programming/profiling-nvidia-gpu-performance.md | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/docs/programming/profiling-nvidia-gpu-performance.md b/docs/programming/profiling-nvidia-gpu-performance.md index 885a77c27..c6923404b 100644 --- a/docs/programming/profiling-nvidia-gpu-performance.md +++ b/docs/programming/profiling-nvidia-gpu-performance.md @@ -390,6 +390,10 @@ Collecting performance data using `ncu` can incur significant runtime overhead. ### Getting Started +```{attention} +`ncu` is not compatible with MIG-enabled GPUs. Ensure you run `ncu` only on GPU nodes without MIG. +``` + To use `ncu`, first load the appropriate CUDA module: ``` @@ -410,8 +414,6 @@ $ ncu --set full --target-processes all ./vectorAdd - `--target-processes all`: Profiles all child processes (useful for multi-threaded applications). -`ncu` is not compatible with MIG-enabled GPUs. Ensure you run `ncu` only on GPU nodes without MIG. -``` ::::{dropdown} Click here to view the full `ncu` report :icon: note From 6c801ce79bf90f5568be88ae95cabddd781f7b48 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 13:29:29 -0600 Subject: [PATCH 09/13] Update docs/software/alphafold.md Co-authored-by: monaghaa <33913248+monaghaa@users.noreply.github.com> --- docs/software/alphafold.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/software/alphafold.md b/docs/software/alphafold.md index dded738e9..8ba82ebfc 100644 --- a/docs/software/alphafold.md +++ b/docs/software/alphafold.md @@ -112,7 +112,7 @@ Loading the AlphaFold 3 module does the following: - redirects temporary files to `/scratch/alpine/$USER` - you can override this path by resetting TMPDIR *after* you load the module: ``` - module load alphafold/3.0.0 + module load alphafold/3.0.3 export TMPDIR= ``` - creates a shortcut to the AlphaFold 3 script so you can run the program with `run_alphafold` From d53a0e36c964fe296846c070185a6756a776617e Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 13:29:38 -0600 Subject: [PATCH 10/13] Update docs/software/alphafold.md Co-authored-by: monaghaa <33913248+monaghaa@users.noreply.github.com> --- docs/software/alphafold.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/software/alphafold.md b/docs/software/alphafold.md index 8ba82ebfc..97aef4f50 100644 --- a/docs/software/alphafold.md +++ b/docs/software/alphafold.md @@ -146,7 +146,7 @@ To better utilize limited GPU resources, these stages can be split using flags: - `--norun_data_pipeline` → Run only the inference step (Stage 2) #### AlphaFold 3 Examples -Example input files and scripts are in `/curc/sw/install/bio/alphafold/3.0.0/examples`. +Example input files and scripts are in `/curc/sw/install/bio/alphafold/3.0.3/examples`. Loading the AlphaFold 3 module stores this path in `AF3_EXAMPLES`: ``` ls $AF3_EXAMPLES From a4b4c7bf4b820bc2de98a723d042c0c1b5c9742f Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 13:29:45 -0600 Subject: [PATCH 11/13] Update docs/software/alphafold.md Co-authored-by: monaghaa <33913248+monaghaa@users.noreply.github.com> --- docs/software/alphafold.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/software/alphafold.md b/docs/software/alphafold.md index 97aef4f50..bf97f0503 100644 --- a/docs/software/alphafold.md +++ b/docs/software/alphafold.md @@ -161,7 +161,7 @@ You can copy the examples folder to a location where you have write permissions ```bash cd /projects/$USER -cp -R /curc/sw/install/bio/alphafold/3.0.0/examples . +cp -R /curc/sw/install/bio/alphafold/3.0.3/examples . cd examples ``` From 57ce6d8228a14483e1d623469fff65567e6a73a8 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 13:29:52 -0600 Subject: [PATCH 12/13] Update docs/software/alphafold.md Co-authored-by: monaghaa <33913248+monaghaa@users.noreply.github.com> --- docs/software/alphafold.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/software/alphafold.md b/docs/software/alphafold.md index bf97f0503..08672be32 100644 --- a/docs/software/alphafold.md +++ b/docs/software/alphafold.md @@ -184,7 +184,7 @@ Path of the script: `$AF3_EXAMPLES/alphafold3_alpine.sh` # Load the AlphaFold 3 module module purge -module load alphafold/3.0.0 +module load alphafold/3.0.3 # Set input JSON, output directory, and model parameter path export INPUT_FILE=$AF3_EXAMPLES/fold_protein_2PV7/alphafold_input.json From 7103dee36914863147b5174ee82f137295608165 Mon Sep 17 00:00:00 2001 From: John Reiland <93950341+jreiland@users.noreply.github.com> Date: Tue, 28 Jul 2026 13:30:04 -0600 Subject: [PATCH 13/13] Update docs/software/alphafold.md Co-authored-by: monaghaa <33913248+monaghaa@users.noreply.github.com> --- docs/software/alphafold.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/software/alphafold.md b/docs/software/alphafold.md index 08672be32..6eb17c514 100644 --- a/docs/software/alphafold.md +++ b/docs/software/alphafold.md @@ -97,7 +97,7 @@ AlphaFold 3 version 3.0.3 is supported on **all NVIDIA GPU types** on Alpine. Load AlphaFold 3 module: ``` -module load alphafold/3.0.0 +module load alphafold/3.0.3 ``` View run options: ```