The RACE I and RACE II architectures are high-performance engines that were designed to be extremely memory efficient; hence, they are the only architectures that access a voxel only once per projection regardless of the degree of over-sampling. Furthermore, it can potentially exhibit frame-to-frame coherence for multiple view positions. Previous DVR solutions rely primarily on one of two forms of acceleration: 1) sample-reduction or 2) large sample-throughput (sample-memory efficiency defined in [3]). We define sample-memory efficiency as the ability to render the dataset as fast as it can be read from the voxel-memory subsystem when one-to-one voxel-to-samples is present and there is no sample reduction. This is the theoretical maximum performance for rendering volumes that exhibit one-to-one sampling and no sample reduction. To date DVR-architectures have only been able to sufficiently utilize one form acceleration. By combining both forms of acceleration, the RACE II architecture is expected to have a two-fold speedup over other solutions. The RACE I architecture will achieve larger average perspective performance than other solutions that have more than double the amount of voxel-bandwidth as the RACE engine.

To achieve a new level of performance RACE implements a truly hybrid algorithm. In a nutshell, it uses object-order control over image-order rendering units. It allows the RACE architectures to trivially handle perspective projections and sample reduction techniques such as Early-Ray Termination (ERT) and Space-leaping. RACE as presented in [4] only utilizes early-ray termination. RACE II supports space-leaping. Based on simulation results the RACE architecture will achieve higher average perspective rendering performance than currently available solutions using anywhere from 33-75% less bandwidth the other solutions (See table below). RACE II renders 256x256x128 volumes at an average of 80-90Hz and 256x256x256 volumes at an average of 44 Hz with worst-case 40 Hz and 20 Hz respectively.

Hardware Description (RACE I/RACE II):

• 4 100 MHz rendering units
• 4 8 MB SDRAMS for volume storage (4-way interleaved)
• 4 2 MB SDRAMS for pixel-ray storage and slice buffers (4-way interleaved)
• 1 Projection and memory control unit.
• 96-bit (12 byte) wide pixel-ray interface to each rendering unit
• 2 byte wide voxel interface to each rendering unit
• Operational frequency is currently 100 MHz
• Worst-case frame rate for one-to-one resampling is 20 Hz
• Average frame rate is 22 Hz for RACE I and 44 Hz for RACE II.
• RACE II requires 4 KByte buffer to store empty-space information in the controller

Current limitations of OTHER approaches

• Sample throughput bottleneck because of the eight-to-one voxel to rendering pipeline topology.
• Lack of scalability (each rendering pipeline needs a copy of the ENTIRE volume memory!)
• Low performance when 20% or more of the dataset contributes to the final rendered image.
• Large sensitivities to dataset type, view position, and classification mapping.
• Requires non-interactive pre-processing overhead
• Performance is inversely related to the size of the interpolation filter

• Lack of support for sample reduction techniques
• Large number of rendering engines required for next generation datasets

The figure above shows that the RACE II architecture achieves the largest amount acceleration when 5-90% dataset contributes to the final image. RACE I achieves better aggregate performance (parallel and perspective projections) than the remaining architectures when 12-100% of the dataset contributes to the final image. Volume Pro excels when 90-100% of the dataset contributes to the final image for parallel projections. The two image-order architectures (plots overlap) will achieve the most acceleration when less than 5% of the dataset contributes to the final image. Since VIZARD II currently only uses Early-Ray termination, it is unlikely that less than 70% of the dataset will contribute to the final image using reasonable opacity thresholds. However, the architecture (called the VG Engine) that utilizes space-leaping may reach as low as10% under certain conditions (i.e., binary opacity classifications of sparse datasets). The RACE engines provides robust acceleration for general purpose Volume Graphics and Volume Visualization since they provide the nearest-to-optimal acceleration for over 85% of the range. As a result of efficient voxel-sample processing, we can achieve comparable or better performance than other solutions using anywhere from 33-75% less voxel-bandwidth!