Massively Parallel Architecture

Higher efficiency can only to a limited degree be offset by increased head count. New tools are needed to increase the throughput. The Headwave solution is the first to utilize the computational power of graphics card for processing resulting in unbeatable performance at a fraction of the cost of proprietary platforms.

Parallelization

In recent years we have seen the oil and gas industry embrace PC based clusters for computational purposes. Based on the Intel or AMD processors, clusters have made new algorithms feasible and reduced the turnaround time significantly - but not enough.

To advance further, Headwave has employed commonly available technology in radical ways. By utilizing the processors ("GPUs") on graphics cards for processing, not only visualization, the Headwave solution is capable of providing results in real time or near real time.

Case : 200 x speed up in data processing

As a case in point for how algorithms on GPUs outperform CPUs, look at these numbers. Headwave contains a highly optimized algorithm for wavelet based compression and decompression of seismic data. Implemented and optimized on the CPU, we are able to compress approximately 10MB/sec. Headwave's implementation on the GPU exploits the parallel capabilities of the GPU, providing more than 2GB/sec - a performance speed up of 200 times.

Other commonly used algorithm in prestack data processing that we have implemented, run anywhere from 10 to 70 times faster on the GPU than on the CPU.

The real fun starts when you look at the aggregate performance numbers with GPUs in clusters, as this explains why Headwave is able to do a lot more in real time. Beyond the use of GPUs, we have also added unique technologies for intra-node data transport, caching mechanisms, remote visualization streaming and more that make sure that every link of our solution is as high-performing as possible.

GPU & CPU

The following illustration shows the GFLOPS (billion floating point operations, a commonly used measure of the computational power of a processor) for the past few years for the CPU and GPU respectively.

This graph tells us two things:

That a typical Intel/AMD CPU is capable of approx. 15 GFLOPS (where as Nvidia or ATI GPU is capable of more than 10 times that of a CPU)

That the annual increase in performance of GPUs far outpace Moore's Law for CPUs.

GPUs in Clusters

The obvious question is - can you put GPUs in clusters? Most clusters deployed in industry as of today do not have GPUs*. With the advent of PCI-Express x16 serverboards, we are now seeing high performance computing cluster nodes capable of housing one or two high-end GPUs. The total compute power of a single node can easily approach 350 GLFOPS - in reality the compute power of more than 10 non-GPU dual-CPU nodes.