About Sparsix
Sparsix Corporation is a computational mathematics company. Our senior mathematicians and computer scientists have spent the last three decades expanding the frontiers of high-performance computational mathematics and machine learning by tackling some of the most demanding applications known.
Working initially in government laboratories, these mathematicians and computer scientists developed ground-breaking methods for modeling the complexities of nuclear detonation simulation. Leveraging this experience in modeling extremely large systems, the team began pursuing commercial opportunities in oil and gas reservoir modeling and industrial process modeling and optimization.
Today Sparsix focuses its efforts on energy-related markets: power transmission and distribution, integration of renewable energy sources, and oil and gas exploration and production.
We believe that our unique skills and experience in modeling and optimizing extremely large systems can bring a new perspective to the computational challenges these markets have faced for many years.
Sparsix is based in San Francisco, California and has R&D operations in Atlanta, Georgia and Moscow, Russia. The R&D team is comprised of over 35 mathematicians and developers, one third of whom are PhDs and the remainder of whom have other advanced degrees. The company maintains strong links with the Russian Academy of Sciences and several US and Russian national laboratories providing access to a large population of specialists in numerous disciplines.
Helping Customers Succeed
Successful companies depend on collaboration to bring together the skills necessary to remain competitive in a global marketplace. The specific value that Sparsix brings to our customers relates not to the domain requirements of any particular vertical market – we rely on our industry partners for that expertise – but rather to the underlying mathematical challenges posed by simulating and optimizing the large systems encountered in these fields.
Computing power continues to grow by leaps and bounds, but performance for computationally complex applications has in many cases not kept pace. Sometimes customers want to do more – more design iterations, higher resolution calculations – whatever it takes to improve quality or reliability or reduce cost.
Other times traditional methods that worked well for smaller problems do not scale well for large problems. For example, when moving from simulating local power grids to regional or national level grids, or moving from simulating 2D systems to complex 3D system, new computational methods are often needed. Yesterday’s solutions no longer provide the performance necessary to ensure the reliability and stability that today’s markets demand.
Success Stories
Sparsix creates new approaches based on our expertise in the following three fundamental areas:
- Fast solution and optimization of linear systems through more efficient mathematics.
At the heart of many of the most challenging problems in science and industry is a linear system that must be solved. Our team specializes in developing custom solvers for specific problems.
LP optimizer for power grid modeling
Working in conjunction with a US national laboratory, Sparsix has developed a linear programming data optimization tool (presolver) that accelerates performance of an existing national-level power system modeling tool by 8X independent of the hardware platform.
- Innovative alternatives to traditional modeling methods based on machine learning and intelligent algorithms.
Machine learning (ML) methods are primarily used in applications where formal, complete algorithms for the solution either do not exist or are too costly and complicated to implement. ML methods are also used in combination with traditional computational methods to provide a new level of adaptation and intelligent behavior to existing solutions.
Sparsix brings decades of experience in implementing real-world ML solutions based on architectures such as neural networks, radial basis and kernel machines, Bayesian networks, probabilistic trees, evolutionary and genetic algorithms, fuzzy logic and neuro-fuzzy machines.
Automatic control of industrial chemical process
Similar in many aspects to managing a power grid, this application required the fast solution of a non-linear optimization algorithm which was further complicated by the uncertainties of various chemical parameters. Implementation of a traditional computational approach would have been extremely difficult and costly. Instead, a neural network trained on historical process data and fed with real-time, measured data was used to control the process in real time.
- Expert application of parallel processing architectures to the solution of linear systems.
The tremendous growth in parallel processing architectures (multi-core, clusters, etc.) presents a wonderful opportunity for seemingly “easy” advances in application performance. However, most current applications are not designed to take advantage of these architectures and many of the most computationally demanding applications are not easily modified to perform certain tasks in parallel. Sparsix has deep experience implementing parallel architectures specifically for solving linear systems.
Parallel iterative solver for complex linear systems
The solution of complex linear systems arising from Helmholtz and Maxwell equations can be very computationally demanding. To address a specific need, the Sparsix team created a custom solver optimized for a clustered architecture that delivered exceptional scalability.