CUDA Spotlight: Paulo Souza




Paulo Souza

By Calisa Cole, posted Nov. 21, 2012
Read more CUDA in Action Spotlights

GPUs for Seismic Imaging

This week's Spotlight is on Paulo Souza of Petrobras, the Brazilian multinational energy company headquartered in Rio de Janeiro.

Paulo has been working for nine years with E&P (Exploration & Production) geophysics software and seismic imaging on HPC clusters.

This interview is part of the CUDA Spotlight Series.

Q & A with Paulo Souza

NVIDIA: Paulo, what do you do at Petrobras?
Paulo: I'm the lead HPC developer of seismic imaging codes. Our group produces state-of-the-art algorithms to generate more accurate subsurface images of earth in order to better find oil.

Our algorithms are used across all of our oil fields, including specific algorithms for ultra-deep areas, where the depth of the water is more than 1500 meters.

NVIDIA: When did you first start using GPU-based systems?
Paulo: I started working with GPUs in 2006, porting seismic imaging applications to CUDA. We achieved gains of up to 10X in performance/price and performance/watt, compared to a traditional multi-million dollar x86 CPU-based cluster.

Since that time, we've invested in five GPU clusters, including the Grifo04 built with Tesla M2050 GPUs. Grifo04 is the fastest supercomputer in Latin America.

NVIDIA: How does GPU computing help you in your work?
Paulo: Finding oil in a very complex geology requires compute-intensive algorithms. Seismic imaging codes use a lot of computational power. The high flops (floating-point operations per second) that we get with GPUs helps us meet that need. 

GPUs get the work done fast, while also allowing us to use more accurate and demanding algorithms. For example, in the Kirchhoff Time Migration algorithm we got a 20X speedup when comparing a GPU node against a CPU node.

We like using CUDA because of the ability to write a "scalar" code that is internally vectorized by the compiler. Our codes are really clean and readable.

The Petrobras Grifo04

The Petrobras Grifo04 with NVIDIA Tesla M2050 GPUs is the fastest supercomputer in Latin America

NVIDIA: What changes do you expect the industry to undergo in the next decade?
Paulo: Power consumption is already a big concern today. Over the next 10 years, I expect smartphone and tablet processors to flood the HPC market. At Petrobras we are already evaluating NVIDIA's Tegra mobile processor for seismic imaging codes.

Bio for Paulo Souza

Paulo Souza is an HPC consultant/researcher at Petrobras. Paulo has been working for 9 years with E&P production geophysics software, seismic imaging on HPC clusters, RTM, One Way Wave Equation, Kirchhoff, multiple architecture optimization (GPGPU, x86, Power, Cell) and cluster deployment.

His professional interests include GPGPU, new architectures for HPC, portable fault tolerant MPI applications and fault tolerant I/O, code optimization and SIMD assembly programming, auto-tuning, vectorization, and performance tuning tools. His education includes a B.Sc. in Computer Engineering from Pontifícia Universidade Católica do Paraná.

Relevant Links
http://www.petrobras.com.br

Contact Info
http://www.linkedin.com/in/paulo7

prps (at) petrobras.com.br