AMD Explores PEPS, a New Positional Encoding Method That Could Improve Neural Texture Compression
AMD has introduced new research at the I3D Symposium under the title “PEPS: Positional Encoding Projected Sampling,” outlining a fresh approach to positional encoding that could make neural texture compression more efficient and memory-friendly.
The idea is especially relevant as modern games continue to demand larger textures, higher-resolution assets, and more GPU memory. With many graphics cards still limited by VRAM capacity, smarter compression techniques could become increasingly important for future real-time rendering.
Neural texture compression is a technique that uses machine learning to represent texture data more compactly. Instead of storing every texture detail in a traditional format, the system trains what are known as Implicit Neural Representations, or INRs.
These INRs learn a coordinate-to-signal function. In simple terms, the model is trained to understand how a specific texture coordinate maps to a color or texture value. Texture coordinates are first projected into a higher-dimensional representation, then passed through a small neural network, usually a multi-layer perceptron. This allows the texture to be reconstructed from a much smaller learned representation.
The result is potentially significant texture compression, which could help reduce memory usage in games and 3D applications.
AMD’s PEPS method aims to improve this process by changing how positional encoding works.
Traditional positional encoding takes low-dimensional coordinates and projects them into a higher-dimensional sine and cosine vector. PEPS builds on this concept by treating each sine and cosine projection as a point on a Lissajous curve. The method then samples the encoder or grid at those projected points.
By doing this, PEPS increases the amount of information available to the Implicit Neural Representation, improving how effectively the model can represent texture data.
In neural texture compression, this can help produce better results with fewer parameters, meaning the model may need less storage while still preserving visual detail. That is the key appeal: reducing memory requirements without dramatically sacrificing texture quality.
However, the trade-off is performance cost.
AMD’s testing showed that generating a 1024×1024 three-channel texture on a 9070 XT took 4.32 ms using the BI-grid baseline. With Grid-PEPS, that increased to 5.47 ms. A more optimized version called Grid-PinkPEPS reduced the time to 4.86 ms.
This slowdown comes from the extra sampling, memory access, and computation required by the PEPS technique. So while the method can improve compression efficiency, it also adds processing overhead.
That balance between memory savings and compute cost will be one of the biggest factors in determining whether techniques like PEPS become useful in real-time applications such as games.
AMD also explored PEPS beyond standard neural texture compression. One promising area is signed distance functions, or SDFs, which are widely used in 3D rendering to represent shapes and surfaces.
SDFs can be extremely memory-hungry because they often require high-resolution grids to accurately represent detailed geometry. This can quickly consume VRAM, especially in complex scenes. Neural compression methods could help reduce that burden by storing the same type of information more efficiently.
In AMD’s testing with the Pitted Stonefish SDF, Grid-PEPS was able to roughly match the Intersection over Union accuracy of non-PEPS methods that used eight times more encoder parameters. Intersection over Union, often shortened to IoU, measures how closely a reconstructed 3D shape matches the original.
That result suggests PEPS could be useful for compact neural representations of 3D geometry, not just texture data.
Still, this does not mean gamers should expect PEPS-powered neural texture compression to appear in Radeon drivers or major game engines anytime soon.
At the moment, neural texture compression remains mostly a research and development topic. There are no widely available games using a complete implementation of this technology, and consumer-facing support is still limited. AMD has also not established a consumer brand name for this type of neural compression work, instead referring to it through general technical terms in its research.
Even so, the research is important. GPU memory pressure is not going away. High-resolution textures, ray tracing, advanced geometry systems, and larger open-world environments continue to push VRAM requirements higher. Meanwhile, many 8GB graphics cards are likely to remain in use for years.
If neural texture compression and related methods mature, they could help developers deliver richer visuals while reducing memory usage. PEPS is not a finished gaming feature yet, but it is another step toward more efficient rendering pipelines and smarter use of GPU resources.
For now, AMD’s PEPS research is best seen as a promising technical experiment. It shows how changes to positional encoding can improve neural representation efficiency, particularly for texture compression and SDF rendering. The challenge will be turning that research into practical tools that work reliably in real-time games without adding too much performance overhead.
