Intel's TSNC neural texture compression can drastically shrink game textures and cut VRAM use

Intel introduced TSNC, a neural texture compression method that builds on BC1 block compression by training on related texture sets and storing a shared latent.

Intel reported internal results where standard BC reached ~4.8× compression, TSNC Variant A exceeded 9×, and Variant B exceeded 18×, with perceptual loss estimates around 5%–7%.

Intel demonstrated TSNC on upcoming Panther Lake integrated graphics with a decompression API and measured a 3.4× speedup using XMX versus an FMA fallback.

Intel plans to release TSNC as a standalone SDK, starting with an alpha later this year and progressing to beta and public releases.