Intel TSNC: The AI Revolutionizing Video Game Compression

Intel TSNC: A Major Breakthrough in Video Game Compression

Intel has recently unveiled its new technology, TSNC or Texture Set Neural Compression. This innovation drastically reduces the size of textures in video games by up to 18 times. The problem it addresses is well-known: as video games strive for ever-increasing realism, they require more and more detailed textures. Each visual element, whether reflections, shadows, or textured surfaces, demands its own layer of information, significantly increasing the size of game files.

This accumulation of data not only affects storage but also puts a strain on video memory, or VRAM. When this memory is saturated, game performance can drop significantly, breaking the player's immersion. In this context, simply increasing hardware power is no longer sufficient, which is where TSNC comes into play.

An Integrated and Flexible Approach

Intel has opted for a solution that seamlessly integrates with existing methods rather than creating an entirely new format. TSNC builds on the widely adopted block compression methods, notably the BC1 format. This means that developers do not need to overhaul their development processes to take advantage of this new technology.

However, Intel is not alone in this space. Nvidia is working on its own neural compression technology, called NTC, and rumors suggest that Sony may also be developing similar solutions for its upcoming console, the PS6.

How TSNC Works

Rather than compressing each texture individually, TSNC employs a holistic approach. It trains a neural network on a set of similar textures, which it then groups into a common space. This data is organized according to several levels of traditional compression.

The reconstruction of the compressed data is handled by another neural network, a three-layer MLP. This system offers great flexibility, allowing developers to choose when to decompress based on their needs: this can occur during game installation, loading times, or even during gameplay. The goal of decompression can vary depending on priorities: reducing download size, limiting bandwidth usage, or easing video memory consumption.

Two Variants for Varied Needs

To address different requirements, Intel has developed two distinct profiles for TSNC. Variant A offers a balance between compression and visual quality. It allows for texture size reduction of up to nine times, with a visual loss of about 5%, measured using Nvidia's FLIP tool. In comparison, traditional methods struggle to achieve a ratio of 4.8x.

Variant B, on the other hand, aims for maximum compression, achieving a spectacular ratio of 18x. This approach results in slightly higher visual degradation, between 6 and 7%, with the appearance of some subtle artifacts. However, these compromises are often acceptable for players who prioritize smoothness on less powerful machines.

Performance and Compatibility

Intel has conducted extensive testing to evaluate the impact of its technology on performance. On the Panther Lake architecture, equipped with an integrated Arc B390 GPU, TSNC leverages XMX cores specifically designed for artificial intelligence. The result is impressive: the system can generate a first pixel in just 0.194 nanoseconds, making latency practically imperceptible to the human eye.

For users with older configurations or competing GPUs, Intel offers an alternative solution. Although this method is about 3.4 times slower, at 0.661 nanoseconds per pixel, it remains fast enough to ensure a smooth gaming experience.