AMD Unveils New Technique to Reduce Graphics Memory Usage
AMD has presented a new technique called "PEPS" that reduces model parameters by 25% while maintaining image quality. The technology can significantly decrease graphics memory requirements.
AMD recently unveiled its "PEPS: Position Encoding Projection Sampling" research at the I3D workshop. This technique focuses on neural texture compression, successfully reducing model parameters by 25% without compromising visual quality.
PEPS aims to optimize graphics memory consumption by utilizing an "implicit neural representation" model. This model learns to map texture coordinates to final signal values. Compared to traditional methods, PEPS involves a more detailed mapping of projections, allowing it to achieve similar results with fewer parameters.
A trade-off for this reduction is an increase in computational time. Tests on the Radeon RX 9070 XT graphics card showed that the PEPS solution took slightly longer to generate a 1024x1024 texture than the baseline method. However, optimized versions of the technique have narrowed this performance gap.
The technology also holds potential for 3D rendering applications, such as compressing memory-intensive Signed Distance Functions (SDF). PEPS has demonstrated its ability to substantially decrease memory requirements while preserving the accuracy of 3D shapes.
While promising, the widespread adoption of PEPS in consumer products may take time. The gaming industry currently lacks broad support for neural texture compression, and AMD is still in the early stages with this technology. However, given the ongoing constraints on graphics memory, such optimization techniques are likely to become increasingly important.