Intel's New Neural Compression Tech Can Work On GPUs Without AI Cores

One concern following NVIDIA's introduction of its ultra-efficient Neural Texture Compression (NTC) was that it would remain proprietary to only NVIDIA's GPUs. Intel has since introduced its own similar technology, dubbed Texture Set Neural Compression (TSNC), which boasts similar benefits, without having a specific AI core requirement or being locked to certain hardware. At least, not TSNC Variant B, which compresses textures by up to 18x at the cost of a 7% loss in visual fidelity, according to Intel's data. In comparison, Variant A compresses by roughly 9x. Variant B is also slower, but the functionality is impressive since it was originally intended for AI acceleration via Intel's XMX engines.


The results of the compression look impressive and are comparable to the original material, even upon close inspection. Intel claims that TSNC is meant for developer use an offer multiple benefits: accelerating install time by reducing file sizes, faster streaming of textures as a game loads, streaming textures during real-time gameplay, and loading textures without holding them in VRAM in an uncompressed state.


The benefits of compression technologies like this are obvious in terms of potential bandwidth and storage savings. In the future, we suspect games will ship with support for both Intel TSNC and NVIDIA NTC, or a similar alternative will find its way into DirectX and Vulcan.


What remains to be seen is how Intel TSNC actually performs in games or higher-fidelity, path traced scenes. NVIDIA was a lot more eager to show off its texture compression with large, complex scenes, while Intel's presentation is more focused on individual models and texture sets. In any case, we look forward to seeing how these neural compression technologies mature and ultimately come to market.