Caustic Graphics
| Company type | Private |
|---|---|
| Industry | Semiconductor |
| Founded | 2006 |
| Founder | Three ex-Apple engineers |
| Fate | Acquired by Imagination Technologies in 2010 |
Area served | Worldwide |
| Products | Real-time ray tracing hardware and software solutions, PCIe cards (R2100, R2500, CausticOne) |
| Parent | Imagination Technologies (from 2010) |
Caustic Graphics was a fabless semiconductor firm dedicated to advancing real-time ray-traced graphics technology designed for broad consumer use. Established in 2006 by three ex-Apple engineers—one of whom contributed to graphics development for the 5th Generation iPod and the original iPhone—the company aimed to optimize GPU hardware to handle ray processing with the same efficiency traditionally reserved for vertex and fragment processing in rasterization.
History
Initially, Caustic concentrated on creating a comprehensive real-time ray tracing solution intended to accelerate professional 3D visualization workflows, with plans to later expand into PC and gaming console sectors. In 2008, Caustic expanded its offerings by acquiring Splutterfish, the developers behind the Brazil R/S production rendering software.
Building on this acquisition, Caustic introduced "Visualiser Plugins," a commercial software suite that combined Brazil R/S rendering technology with Caustic’s hardware. These plugins enabled both interactive and offline ray-traced rendering within prominent 3D design applications including Autodesk 3DS Max, Maya, McNeel Rhino, and SketchUp.
In 2010, Imagination Technologies acquired Caustic Graphics, integrating their advanced cinema-quality ray tracing technology into Imagination’s graphics intellectual property portfolio.
Hardware
The specialized Caustic ray tracing hardware was designed to complement standard rasterisation GPUs by incorporating dedicated features that enhanced ray-tracing performance:
- Coherence gathering: Allowed the GPU to identify and cluster work units that requested shared resources such as memory locations or shader programs.
- Fixed-function triangle intersection units: Performed ray-triangle edge intersections, outputting barycentric coordinates along with hit details.
- Fixed-function bounding volume intersection units: Conducted intersections between rays and AABB volumes, returning relevant hit information.
- Framebuffer accumulator: Applied color contributions to pixels in the frame buffer based on ALU instructions.
- Scene hierarchy generator: Converted geometry lists into a bounding volume hierarchy—a directed acyclic graph representing 3D bounding volumes enveloping scene geometry.
Coherence gathering
Caustic’s architecture enabled parallel execution at the granularity of individual rays rather than pixels. This allowed the hardware to coherently group rays performing similar operations—such as accessing memory within the scene acceleration structure or running identical material shader programs. This method significantly boosted efficiency during ray-tracing tasks.
Streaming bounding volume hierarchy generator
Caustic engineered a hardware pipeline capable of constructing a bounding volume hierarchy in a bottom-up fashion by directly processing assembled primitives post-vertex shading. This pipeline streamed out a graph of AABB nodes to DRAM in a single pass. It utilized a conservative triangle voxelizer that generated spatial addresses within a cache-like system to cluster triangles and AABBs spatially, relying heavily on spatial coherence in triangle submission. When coherence was lacking, it resulted in premature cache evictions, overlapping bounding volumes, and increased overhead during ray traversal. The hardware also enhanced voxelization rates for elongated, narrow triangles to optimize bounding accuracy and improve rendering efficiency for off-axis geometry.
This component, known as the SHG, was incorporated into Imagination’s PowerVR 6XT GR6500 GPU. According to Caustic’s testing, the SHG achieved throughput rates of up to 100 million dynamic triangles per second on silicon clocked at 600MHz.
CausticGL/OpenRL API
The OpenRL API, previously known as CausticGL, was developed atop OpenGL ES 2.0 and enhanced both the host interface and GLSL to incorporate ray tracing capabilities. Caustic delivered optimized implementations tailored for Intel CPUs utilizing SSE and AVX, alongside support for GPUs compatible with OpenCL and CUDA-enabled NVIDIA hardware.
OpenRL SDK
The OpenRL API was distributed as a complimentary SDK compatible with Intel processors, OpenCL and CUDA GPUs, as well as Caustic's proprietary hardware. Applications developed with this SDK could automatically identify and leverage the most efficient hardware available.
Included within the SDK were sample applications, a "Profiler" tool designed for comprehensive examination of the ray tracing "tree" structure within individual pixels, and "StatsPlotter," which revealed internal performance metrics to facilitate debugging and performance tuning.
PowerVR Brazil SDK
The PowerVR Brazil SDK (also known as Resin) functioned as a software abstraction utilizing the OpenRL API to enable features such as physically based rendering, programmable shader support, dynamic geometry handling, and cross-platform hardware compatibility. This technology stemmed from the Brazil R/S team acquired by Caustic Graphics.
PowerVR Wizard / PLATO
Following its acquisition by Imagination Technologies, Caustic’s innovations—including ray tracing coherence collection, fixed-function geometry testing units, and streaming BVH generation—were incorporated into the Series 6 PowerVR GPU architecture. This integration enabled GLSL shaders, which had previously been executed on CPUs, to run directly on-chip through existing PowerVR shading clusters while benefiting from Caustic’s ray reordering techniques.
The GPU retained its tile-based deferred rasteriser (TBDR) architecture, facilitating hybrid rendering by allowing triangle fragment shaders to dispatch rays into the ray tracing pipeline seamlessly.
Imagination Technologies produced test silicon of the PowerVR 6XT GR6500 ray tracing GPU on a PCIe card named PLATO. This prototype was publicly showcased in 2016, boasting memory bandwidth and power consumption levels comparable to those found in mid-tier smartphones.
PCIe cards
Two PCIe card models incorporated the Caustic RT2 ASIC, fabricated using a 90nm TSMC process. The R2100 model housed a single RT2 chip paired with 4GB of DDR2 memory and had a maximum power draw of 40W. Meanwhile, the R2500 featured dual RT2 chips alongside 16GB of memory, with peak power consumption rated at 60W. Both cards entered the commercial market in 2013.
The CausticOne PCIe card utilized two FPGA-based RTU units operating at 100MHz with single-channel DDR2 memory allocated for scene storage; it interfaced through a 4-lane PCIe 1.0 connection. Demonstrated in 2009, this card experienced limited commercial success during that year.