Nvidia’s new texture compression demo would possibly make it easier to overlook about low VRAM in its GPUs

In the course of the preliminary showcase of the RTX 50 sequence GPUs, individuals criticized low VRAM specs, particularly in the RTX 5070 which only features 12GB. And the criticism wasn’t flawed – fashionable video games have grow to be more and more memory-hungry, and 12GB doesn’t really feel like greater than sufficient in some instances. Nevertheless, the confusion was quickly cleared up when Nvidia’s CEO, Jensen Huang, explained that the new AI technology in these next-gen GPUs is not only 40% faster but also uses 30% less VRAM, making it extra memory-efficient.

Now, Nvidia’s RTX Neural Texture Compression (NTC) has lastly been benchmarked, giving RTX 50 sequence house owners extra reminiscence effectivity to stay up for. And first impressions? It does look fairly promising. This comes from Compusemble, who benchmarked Nvidia’s new reminiscence compression tech on an RTX 4090 at 1440p and 4K decision. The outcomes had been spectacular, as Nvidia’s neural compression know-how considerably diminished a 3D utility’s texture reminiscence footprint. However there is a catch.

96% discount in reminiscence texture dimension, however at the price of efficiency

Compusemble examined NTC in two modes: “NTC transcoded to BCn” and “Inference on Pattern.” There’s much more to grasp about these however to place it merely, the previous converts textures right into a extra environment friendly format when loading a sport, whereas the latter solely hundreds the elements of a texture which can be truly wanted for what’s on display screen.

At 1440p with DLSS upscaling enabled, the “NTC transcoded to BCn” mode diminished texture reminiscence by 64%, dropping it from 272MB to 98MB. The “NTC inference on pattern” mode diminished the feel dimension even additional to simply 11.37MB – a 95.8% discount in reminiscence utilization in comparison with non-neural compression and an 88% discount in comparison with the earlier neural compression mode. Nevertheless, this did come at a efficiency value. Whereas the “NTC transcoded to BCn” mode had a negligible influence on common FPS in comparison with NTC off, “NTC inference on pattern” mode took the largest hit, dropping from the mid-1,600 FPS vary to the mid-1,500s, with 1% lows falling to round 840 FPS.

As for 1440p with TAA anti-aliasing as an alternative of DLSS upscaling, reminiscence capability discount remained the identical, however the GPU’s efficiency habits modified. All three modes ran considerably sooner than with DLSS. 4K efficiency was additionally fairly fascinating, as native 4K TAA benchmarks confirmed a median FPS increase over DLSS-enabled efficiency. There’s much more that Compusemble covers of their video, together with cooperative vector testing and different particulars, so we extremely advocate checking it out.

A variety of GPUs would possibly be capable of profit from this

The important thing takeaway right here is that Nvidia’s plan to compress textures “by another 5X” with a brand new shader processor in Blackwell GPUs seems to be the principle purpose, and this demo has given us our first have a look at it. Whereas all of this does come at a efficiency value, the reminiscence effectivity positive factors will not be one thing to ignore for the sake of some misplaced frames. The DLSS vs. native decision efficiency distinction means that the tensor cores dealing with RTX NTC are being closely taxed, probably to the purpose of bottlenecking the shader cores. This may clarify the numerous body charge improve at native decision; in any other case, we’d count on DLSS mode to run at a better body charge than the native 4K TAA benchmarks.

That stated, the know-how continues to be in beta, and there’s no official launch date but. Nevertheless, one fascinating element is that based on Nvidia’s GitHub web page for RTX NTC, the minimal GPU requirement is an RTX 20-series GPU. But, the know-how has additionally been validated to work on GTX 10-series GPUs, AMD Radeon RX 6000-series GPUs, and Arc A-series GPUs. This means we might see the know-how grow to be mainstream on non-RTX GPUs and probably even consoles within the close to future.