During the initial showcase of the RTX 50 series GPUs, people criticized low VRAM specs, especially in the RTX 5070 which only features 12GB. And the criticism wasn't wrong – modern games have become increasingly memory-hungry, and 12GB doesn’t feel like more than enough in some cases. However, the confusion was soon 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 more memory-efficient.

Now, Nvidia's RTX Neural Texture Compression (NTC) has finally been benchmarked, giving RTX 50 series owners more memory efficiency to look forward to. And first impressions? It does look pretty promising. This comes from Compusemble, who benchmarked Nvidia's new memory compression tech on an RTX 4090 at 1440p and 4K resolution. The results were impressive, as Nvidia's neural compression technology significantly reduced a 3D application’s texture memory footprint. But there's a catch.

96% reduction in memory texture size, but at the cost of performance

Compusemble tested NTC in two modes: “NTC transcoded to BCn” and “Inference on Sample.” There's a lot more to understand about these but to put it simply, the former converts textures into a more efficient format when loading a game, while the latter only loads the parts of a texture that are actually needed for what's on screen.

At 1440p with DLSS upscaling enabled, the “NTC transcoded to BCn” mode reduced texture memory by 64%, dropping it from 272MB to 98MB. The “NTC inference on sample” mode reduced the texture size even further to just 11.37MB – a 95.8% reduction in memory utilization compared to non-neural compression and an 88% reduction compared to the previous neural compression mode. However, this did come at a performance cost. While the “NTC transcoded to BCn” mode had a negligible impact on average FPS compared to NTC off, “NTC inference on sample” mode took the biggest hit, dropping from the mid-1,600 FPS range to the mid-1,500s, with 1% lows falling to around 840 FPS.

As for 1440p with TAA anti-aliasing instead of DLSS upscaling, memory capacity reduction remained the same, but the GPU's performance behavior changed. All three modes ran significantly faster than with DLSS. 4K performance was also quite interesting, as native 4K TAA benchmarks showed an average FPS boost over DLSS-enabled performance. There’s a lot more that Compusemble covers in their video, including cooperative vector testing and other details, so we highly recommend checking it out.

A lot of GPUs might be able to benefit from this

The key takeaway here is that Nvidia’s plan to compress textures “by another 5X” with a new shader processor in Blackwell GPUs appears to be the main goal, and this demo has given us our first look at it. While all of this does come at a performance cost, the memory efficiency gains are not something to disregard for the sake of a few lost frames. The DLSS vs. native resolution performance difference suggests that the tensor cores handling RTX NTC are being heavily taxed, potentially to the point of bottlenecking the shader cores. This would explain the significant frame rate increase at native resolution; otherwise, we would expect DLSS mode to run at a higher frame rate than the native 4K TAA benchmarks.

That said, the technology is still in beta, and there is no official release date yet. However, one interesting detail is that according to Nvidia’s GitHub page for RTX NTC, the minimum GPU requirement is an RTX 20-series GPU. Yet, the technology has also been validated to work on GTX 10-series GPUs, AMD Radeon RX 6000-series GPUs, and Arc A-series GPUs. This suggests we could see the technology become mainstream on non-RTX GPUs and possibly even consoles in the near future.

About the Author

Hassam Nasir

Hassam boasts over seven years of professional experience as a dedicated PC hardware reviewer and writer.