AMD’s PEPS Research Cuts Neural Texture Compression Size by 25%
AMD has detailed a new piece of graphics research called PEPS, or “Positional Encoding Projected Sampling,” that shrinks the neural networks used to compress game textures by roughly a quarter without a meaningful drop in visual fidelity. The work was presented at the I3D 2026 symposium, the same venue AMD used last year to showcase related neural rendering research, according to a paper published by the company’s GPUOpen research team.
Neural Texture Compression works by training what are known as INRs, or “Implicit Neural Representations,” to learn coordinate-to-signal functions. By projecting texture coordinates into a higher-dimensional embedding and feeding this information to a multi-layer perceptron, it’s possible to represent and compress textures significantly. PEPS is aimed squarely at making that underlying process leaner.
How PEPS Reworks Positional Encoding
PEPS introduces a new method to improve the efficiency of this process by changing how positional encoding is used, given that positional encoding typically projects low-dimensional coordinates into a higher-dimensional sine/cosine vector. PEPS builds upon this by treating each sine/cosine projection as a point on a Lissajous curve, then sampling the encoder/grid at those projected points, a shift AMD’s researchers say lets the network pack more information into fewer parameters.
The technical paper behind the research, published on arXiv, frames the underlying problem in blunt terms: textures and materials are the backbone of modern 3D engines and physically based rendering used in applications such as games, where many 4K textures must be loaded at the same time, creating a memory bottleneck, which is why neural texture compression has gained interest. Shrinking the size of the neural models themselves, rather than just the textures they represent, is the specific gap PEPS is trying to close.
Encoder Parameters Cut Eightfold in Geometry Tests
Beyond flat textures, AMD’s researchers also tested the method on signed distance fields (SDFs), a format used for representing 3D geometry that typically demands high-resolution grids. According to figures reported by Chinese outlet IT之家, in tests on the “Pitted Stonefish” SDF benchmark, the PEPS-enhanced Grid-PEPS model matched the reconstruction accuracy of standard methods while using only an eighth of the encoder parameters, a similar finding independently described by Wccftech’s read of the same research.
When testing on the Pitted Stonefish SDF, Grid-PEPS was able to roughly match the IoU (Intersection Over Union, or how closely the reconstructed 3D shape overlaps the original) of non-PEPS methods with 8x more encoder parameters. For a format notorious for eating into VRAM budgets, that kind of parameter reduction is the whole point of the exercise.
RX 9070 XT Benchmarks Reveal a Compute Trade-off
Smaller models aren’t free, however. IT之家 reported that AMD benchmarked the technique on a Radeon RX 9070 XT, and rendering a 1024×1024, three-channel texture took 4.32ms using the standard BI-grid baseline compared with 5.47ms for the new Grid-PEPS approach — extra time the outlet attributed to the additional sampling steps and memory accesses PEPS requires. A refined variant, called Grid-PinkPEPS, narrowed that gap by bringing the render time down to 4.86ms, per the same report.
In other words, AMD’s research is explicitly trading a small amount of GPU compute time for a meaningfully smaller memory footprint — a trade-off that could matter more on cards where VRAM, not raw shader throughput, is the bottleneck.
No Games Currently Ship With Full Neural Texture Compression
None of this is close to landing in a shipping game yet. Currently, only NVIDIA has any kind of publicly available toolkits/demos for Neural Texture Compression, and there isn’t a single game out there with a full NTC implementation. AMD’s own effort remains earlier-stage still; per IT之家’s reporting, the company has not given the underlying technology a formal commercial brand name, with its research papers continuing to use generic technical terms like PEPS and NTC rather than a marketed feature.
For players in New Zealand and Australia, where imported GPUs carry a persistent price premium and 8GB cards remain common at the budget and mid-range tiers, any technology that squeezes more texture detail out of limited VRAM is worth watching — even years before it reaches a retail driver. Neural texture compression, whether from AMD or Nvidia, would matter most on exactly the class of card that’s hardest to upgrade out of locally: the sub-$500 GPU that has to last a few console generations’ worth of increasingly VRAM-hungry releases.
Read also: SK Hynix ADR Surges 13% on Wall Street Debut Amid AI Memory Boom
AMD’s GPUOpen team has a track record of publishing this kind of foundational research well ahead of any product announcement, as it did with its Neural Texture Block Compression work in 2024. Whether PEPS ends up folded into a future FidelityFX feature, a driver-level tool, or simply informs how AMD approaches VRAM management on upcoming Radeon hardware is not yet clear, and the company has given no timeline for turning the research into something developers can actually ship.
Join the Conversation