Intel TSNC Hits 18x Neural Texture Compression
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Intel detailed TSNC (Texture Set Neural Compression), rebuilt from its GDC 2025 research demo into a standalone SDK with a decompression API that compiles to C, C++, or HLSL.
- TSNC ships in two tiers — Variant A with up to 9x compression and Variant B with up to 18x — compared to roughly 4.8x for standard BC compression in Intel's own tests.
- The TSNC architecture stores latent data in a four-layer BC1-based feature pyramid and reconstructs texture channels through a three-layer MLP decoder.
- Perceptual analysis using NVIDIA's FLIP tool estimated roughly 5% perceptual loss for Variant A and 6–7% for Variant B, with Variant B showing visible BC1 block artifacts in normals and ARM data.
- On Panther Lake B390 integrated graphics, Intel's XMX-accelerated linear algebra path hit 0.194 ns/pixel versus 0.661 ns/pixel for the FMA fallback — about a 3.4x speedup.
- An alpha SDK is planned for later this year, followed by beta and a public release, with TSNC positioned as a runtime path for Intel GPUs with XMX support.
- Intel's effort follows NVIDIA's own neural texture compression demo at GDC, which claimed up to 85% compression, putting both GPU makers on parallel neural codec tracks.
Why it matters: Intel's TSNC SDK ships later this year with up to 18x texture compression, putting Intel and NVIDIA on parallel neural-codec tracks against legacy BC1. The 6–7% perceptual loss at the highest tier forces developers to choose between quality and footprint, while giving Intel a software story to make its GPUs more competitive for game workloads.
Ask SkimNews
