Advanced Rendering Technology The AR250 A New Architecture for Ray Traced Rendering

Advanced Rendering Technology SIGGRAPH ‘99 – Slides Architecture overview AR250 Rendering Processor RenderDrive Network Appliance How and why it works Problems with Parallelism The ART Model Main AR250 Data Flows AR250 Statistics Analysis Performance Functionality Rendering Speed

Advanced Rendering Technology AR250 Rendering Processor Host Bus Interface Host Bus Interface Ray Tracing Engine Ray Tracing Engine Control Unit Control Unit Memory Manager Memory Manager On Chip Data Cache On Chip Data Cache Host System PCI Interface Additional AR250 processors Local DRAM

Advanced Rendering Technology RenderDrive Network Appliance CPU / Motherboard Flash memory boot disk AR250s x 16 Embedded OS RenderMan compliant renderer Compressed RenderMan protocol HTML job control TCP / IP LAN RenderDrive Web Server 3D Workstation RenderPipe Internet Images 3D Application RenderGate

Advanced Rendering Technology Problems with Parallelism Data distribution Load balancing Scalability of calculation Complexity

Advanced Rendering Technology The ART Model Hardware intersection pipeline Ray-parallel data distribution Broadcast parallelism of geometry Hierarchical geometry Distributed concurrent shading Vector parallel programmable shading acceleration

Advanced Rendering Technology Main AR250 Data Flows Host bus interface Pipeline-parallel ray geometry engine (permutes multiple geometry and rays) Pipeline-parallel ray geometry engine (permutes multiple geometry and rays) Programmable shading co-processor Programmable shading co-processor On-chip data caches On-chip data caches Host system interface Geometry Rays Ray generation & control (CPU) Ray generation & control (CPU) Geometry and shaders (broadcast) Pixels (sequential read-back) Additional AR250 processors Intersections Geometry & shaders Shader ops. Pixels Fine grain distribution of rays between AR250s for load balancing Hierarchical geometry Programmable shaders DRAM

Advanced Rendering Technology AR250 Statistics 0.35um drawn, LSI Logic G10 silicon process 650K gates, 106mm 2 die Custom RISC processor core 32, single-stage, 32 bit IEEE compatible floating-point units Multi-dimensional noise, square root and trig. functions 50 MHz operation

Advanced Rendering Technology Analysis Load balancing through fine-grain parallelism Efficient data distribution by caching rays locally and broadcasting geometry Scalability by performing whole rendering calculation, including shading. Low complexity programming model by support for programmable shading in silicon

Advanced Rendering Technology Functionality Current Full shading model Full ray-traced: Camera motion blur Depth of field Area lights Volumes Planned Global illumination Completion of RenderMan support

Advanced Rendering Technology Rendering Speed Hebe mirror (1322 x 2000) RenderDrive (16 x AR250) 0 hrs 13 min 2 sec BMRT (PII - 333) 22 hrs 8 min Speed  C x P 0.82