Zibra AI

ZibraXYZ provides a unified data layer for handling volumetric simulation outputs and delivering them directly to GPUs for machine‑learning workloads. It compresses raw CFD, fluid‑dynamics, climate, structural‑mechanics, electromagnetics and other multi‑physics datasets, then streams the compressed tensors to GPU memory without intermediate CPU staging, aiming to remove storage, I/O and format‑fragmentation bottlenecks.

The system supports a wide range of simulation formats—including structured and adaptive grids, unstructured meshes, surface meshes and custom data—through adaptive encoding that achieves roughly 30× compression for structured grids and 15× for unstructured meshes, with reported throughput of up to 500 GB/s. Decompression occurs natively on the GPU, allowing training pipelines to keep GPUs occupied rather than idle waiting for data.

Targeted at research and engineering teams that run large‑scale physics simulations, ZibraXYZ consolidates the storage, conversion and delivery steps into a single pipeline, enabling faster training of surrogate or foundation models across domains while simplifying collaboration on multi‑terabyte datasets.