At CES 2025, Nvidia introduced so many new issues that it was considerably exhausting to determine simply what was actually value taking note of. Whereas the likes of the RTX 5090 and its monumental price ticket have been grabbing all the headlines, one new piece of tech sat to at least one facet with a number of promise however no game to showcase it. Nonetheless, Nvidia has now launched a beta software program toolkit for its RTX Neural Texture Compression (RTXNTC) system, and after enjoying round with it for an hour or two, I am much more impressed with this than any hulking GPU.
At the second, all textures in video games are compressed into a typical format, to save lots of on space for storing and obtain necessities, and then decompressed when utilized in rendering. It could actually’t have escaped your discover, although, that right this moment’s large 3D video games are…nicely…large and 100 GB or extra is not uncommon.
RTXNTC works like this: The unique textures are pre-converted into an array of weights for a small neural community. When the game’s engine points directions to the GPU to use these textures to an object, the graphics processor samples them. Then, the aforementioned neural community (aka decoding) reconstructs what the texture seems like at the pattern level.
The system can solely produce a single unfiltered texel so for the pattern demonstration, RTX Texture Filtering (additionally known as stochastic texture filtering) is used to interpolate different texels.
Nvidia describes the entire factor utilizing the time period ‘Inference on Pattern,’ and the results are impressive, to say the least. With none type of compression, the texture reminiscence footprint in the demo is 272 MB. With RTXNTC in full swing, that reduces to a mere 11.37 MB.
Picture 1 of 3
The entire technique of sampling and decoding is fairly quick. It isn’t fairly as quick as regular texture sampling and filtering, although. At 1080p, the non-NTC setup runs at 2,466 fps however this drops to 2,088 fps with Interfence on Pattern. Stepping the decision as much as 4K the efficiency figures are 930 and 760 fps, respectively. In different phrases, RTXNTC incurs a body fee penalty of 15% at 1080p and 18% at 4K—for a 96% discount in texture reminiscence.
These body charges have been achieved utilizing an RTX 4080 Tremendous, and lower-tier or older RTX graphics playing cards are more likely to see a bigger efficiency drop. For that type of {hardware}, Nvidia suggests utilizing ‘Inference on load’ (NTC Transcoded to BCn in the demo) the place the pre-compressed NTC textures are decompressed as the game (or demo, on this case) is loaded. They’re then transcoded in an ordinary BCn block compression format, to be sampled and filtered as regular.
The feel reminiscence discount is not as impressive however the efficiency hit is not anyplace close to as huge as with Interfence on Pattern. At 1080p, 2,444 fps it is nearly as quick as an ordinary texture pattern and filtering, and the texture footprint is simply 98 MB. That is a 64% discount over the uncompressed format.
All of this may be for nothing if the texture reconstruction was garbage however as you’ll be able to see in the gallery under, RTXNTC generates texels that look nearly equivalent to the originals. Even Inference on Load seems the similar.
Picture 1 of 3
After all, this is an indication and a easy beta one at that, and it is not even remotely like Alan Wake 2, by way of texture decision and atmosphere complexity. RTXNTC is not appropriate for each texture, both, being designed to be utilized to ‘physically-based rendering (PBR) supplies’ quite than a single, primary texture.
And it additionally requires cooperative vector assist to work as rapidly as this and that is primarily restricted to RTX 40 or 50 sequence graphics playing cards. A cynical PC fanatic would possibly be tempted to say that Nvidia solely developed this method to justify equipping its desktop GPUs with much less VRAM than the competitors, too.
However the tech itself clearly has a number of potential and it is potential that AMD and Intel are engaged on growing their very own methods that obtain the similar end result. Whereas three proprietary algorithms for decreasing texture reminiscence footprints aren’t what anybody needs to see, if builders present sufficient curiosity in utilizing them, then one among them (or an amalgamation of all three) would possibly find yourself being an ordinary side of DirectX and Vulkan.
That might be the finest end result for everybody, so it is value keeping track of AI-based texture compression as a result of identical to with Nvidia’s different first-to-market applied sciences (e.g. ray tracing acceleration, AI upscaling), the business finally adapts them as being the norm. I do not suppose this implies we’ll see a 20 GB model of Baldur’s Gate 3 any time quickly however the future actually seems rather a lot smaller.
If you wish to check out the NTC demo by yourself RTX graphics card, you’ll be able to seize it instantly from right here. Extract the zip and then head to the bin/windows-x64 folder, then simply click on on the ntc-renderer file.
Source link
#Ive #testing #Nvidias #Neural #Texture #Compression #toolkit #impressive #results #good #news #game #install #sizes
