Benchmarks
This page contains benchmarking results for the Kelvin2 system.
Intel Data Center GPU Max 1100
This section shows the results of a performance comparison between Intel and NVidia GPUs on the Kelvin2 HPC system. The training scripts may be found here
Benchmark 1: Matrix Multiply (PyTorch)
Method
- Multiply two 8192x8192 matrices using PyTorch
- Run multiple iterations and measure execution time
- Convert throughput to TFLOPS
Results
Benchmark 2: Convolutional Neural Network (PyTorch)
Method
- Run a standard ResNet‑50 training loop on each GPU
- Use synthetic image data created directly on the GPU
- Measure end‑to‑end training throughput (images/sec)
- Use automatic mixed precision (AMP)
Results
Benchmark 3: Molecular Dynamics Simulation (GROMACS)
Method
- Use a standard benchmark from https://www.mpinat.mpg.de/grubmueller/bench
- BenchPEP-h: 12M atoms, Peptides in water, 2 fs time step, h-bonds constrained
Results
AMD Mi300x
This section shows the results of a performance comparison between Intel and NVidia GPUs on the Kelvin2 HPC system.
Benchmark 1: Molecular Dynamics Simulation (GROMACS)
Method
- Use a standard benchmark from https://www.mpinat.mpg.de/grubmueller/bench
- BenchPEP-h: 12M atoms, Peptides in water, 2 fs time step, h-bonds constrained
Results
- Notice the speedup from using Gromac's threaded MPI to the full MPI implementation
Benchmark 2: NAMD
Method
- Use a standard benchmark from https://www.ks.uiuc.edu/Research/namd/benchmarks/previous_benchmarks.html
- STMV 1.06M atom benchmark
Results
