Virtual Molecule Screening (VMS) is a computational technique used in drug discovery that uses machine learning to predict if a chemical compound is likely to bind to a drug target.
This repo contains several implementations of VMS generated from a Matrix Factorization model built by SMURFF
- vms/pure_c: pure C with OmpSs, OpenMP, MPI, GASPI, and ArgoDSM
- vms/af_py: Python ArrayFire implementation for CUDA, OpenCL and MKL
- vms/af_cpp: C++ ArrayFire implementation for CUDA, OpenCL and MKL
- vms/openacc: OpenACC implementation, with and witouth OmpSs
- vms/smurfference: Python NumPy, Tensorflow implementation
- vms/fpga: Xilinx FPGA implementation, with OpenCL and with OmpSs@FPGA
The paper Virtual Screening on FPGA: Performance and Energy versus Effort contains a comparison of VMS on FPGA, CPU and GPU.
Comparison of energy performance of the VMS application implemented on an Nvidia A100 GPU, an Intel Skylake CPU (24 cores @ 2.7Ghz) and a Xilinx Alveo U200 FPGA:
| CPU | GPU | FPGA | |
|---|---|---|---|
| Peak Performance (GF/s) | 3072 | 19500 | 684 |
| Achieved Performance (GF/s) | 402 | 3265 | 260 |
| % of Peak Performance | 13% | 17% | 38% |
| Measured Power Drain (Watt) | 205 | 200 | 37 |
| Energy Efficiency (GF/s/Watt) | 1.8 | 10 | 3 |
1 billion compounds, on a A100 GPU, would take 33 minutes:
N (#samples) 100
K (latent dim) 64
F (#features/compound) 500
P (#proteins) 500
Flops/compound 6400000
achieved gflops/second 3265
compounds/second 510156
mega-compounds/second 0.510
seconds/billion compounds 1960
minutes/billion compounds 33