Engineers are often faced with a variety of simulation prep tasks:
Fit or adapt a structured grid
Define an implicit distance field
Cast and distribute point clouds
Mesh complex geometries
Transform, scale, rotate, translate, flip
Import / export neutral data formats
Common FEA and CFD require a variety of techniques!
How can one yield high performance across the whole process when you're limited by your slowest sub-process? (See: Amdahl's Law)
A unified platform leveraging functional polymorphism and parallel processing brings advanced high-performance computing (HPC) to the desktop and cloud. Our ICCFD10 paper introduces geometry type specialization (grid, mesh, tree, etc) with characteristic uses, forming a complete basis for emerging physical modeling challenges.
Unification requires an underlying base interface, defining common functions across all geometry types. This allows Xplicit-defined and user-defined generic algorithms to be cross-compatible and optimal to later bindings. These modular systems and algorithms yield more than 50x run-time performance benefit and more than 25x faster on GPU in preliminary bench-marking. Lossless protocol buffer compression further minimizes size of *.xcg files and wire transmissions and allows for parallel IO.
Structured grids are known for their efficiency but are topologically constrained in local refinement. Unstructured meshes excel in resolving details but do so at steep memory and computational cost. Large, complex problems have no single good discretization, so they must be decomposed into more thoughtful sub-domains (systems with geometry) each optimized for their physics and integrated at a higher-level (with coupling algorithm).
An example with 1GB RAM: a mesh with 1M elements utilizes the same resources (memory footprint and interconnect traffic) as an optimized 50M element grid. For many configurations, using structured geometries shifts the bottleneck back onto the processor (rather than IO), but remain unsuitable for many real-world scenarios. A hybrid (multi-domain) approach enables more efficient numerical decomposition and enables simulations to scale naturally on memory limited devices such as GPUs.
In real-world practice, XCOMPUTE can yield more than 10x size-up and 10x speed-up compared to current technologies on similar hardware. This means your business can get the job done faster, better, and for the same or less cost.
Learn more about our approach to computational geometries: