AI instead of pricey RAM? What neural rendering really saves in Xbox and PS6
Microsoft is pitching neural rendering as a fix for expensive memory chips. But the buzzword covers very different techniques, and almost all of them save compute time, not memory. We compare Nvidia's 85 percent demo with the only game so far that actually ships something like it.

What this is about
On November 15, 2026, to mark the Xbox's 25th anniversary, Microsoft published a magazine interview with Jason Ronald. He's in charge of the next console generation at Xbox, codenamed Project Helix. The key technology he points to is neural rendering:
“It's a completely new way to generate an image without needing more compute, additional memory and things like that.”
The timing makes sense. According to Ronald, the “current component crisis” is driving up costs for “almost anything you can plug into a wall.” Back in June, when it raised Xbox Series prices, Microsoft wrote that prices for RAM and storage in consoles had risen more than two-and-a-half-fold, and that it expected another doubling by fall 2027. Meanwhile, Sony and AMD are touting “Neural Arrays” and “Universal Compression,” and Nvidia is promising 85 percent less VRAM.
So can AI replace expensive memory chips? Here's our thesis up front: most of what goes by “neural rendering” saves compute work on pixels and bandwidth, meaning the amount of data flowing between memory and the GPU each second. Only one single technique actually saves memory capacity, and so far it runs almost exclusively in demos.
One buzzword, many techniques
“Neural rendering” just means that small neural networks take over parts of the image computation. What they save varies widely:
| Technique | What it does | Saves compute time? | Saves VRAM? |
|---|---|---|---|
| AI upscaling (DLSS, FSR 4, PSSR) | upscales an image from a lower resolution | yes | barely, the network itself needs memory |
| Frame generation | inserts artificial in-between frames | yes, per displayed frame | no, it costs extra memory |
| Ray Reconstruction | denoises ray-traced images with AI | indirectly | no |
| Neural Radiance Cache | a small network learns during gameplay how light is distributed, saving rays | yes | no |
| Neural texture compression | stores surface images as compact data plus a tiny network | no, it costs compute time | yes, but only in one mode |
| Universal Compression (AMD/Sony) | hardware compresses all data on its way to the GPU | – | mainly bandwidth |
For frame generation, the memory cost is documented. With DLSS 4, Nvidia advertised that the new frame generation model needs 30 percent less VRAM than DLSS 3. If there were no memory cost, there'd be nothing to reduce. Apex Compute lists 214 MB for its new frame generation, ANR FG, and 562 MB for Nvidia's DLSS FG (both vendor measurements on an RTX 5090 at 4K). Tools like Veyra, which upscale PS5 streams to 120 frames per second, don't save the console any memory either. They only kick in once the frame is finished.
Then there's infrastructure that saves nothing on its own. It's what makes neural networks fast enough to run in games in the first place. According to heise, Sony and AMD's Neural Arrays link several compute units within a shader engine so they can tackle AI tasks together. On the software side, shaders, the small programs that compute each pixel, need direct access to the GPU's matrix units. For that, Microsoft introduced Cooperative Vectors in 2025 as part of Shader Model 6.9. In September 2025, however, Microsoft announced it would discontinue that spec and fold it into a broader linear algebra design that “will not ship in Shader Model 6.9.” That design is called LinAlg and has been available as a preview in Shader Model 6.10 since April 2026. So under DirectX, the path for neural shaders still isn't cleared for shipping games.
How neural texture compression saves memory, and when it doesn't
Textures are the images wrapped around 3D objects: color, surface relief, roughness and more. Today they're mostly stored in block compression (BCn), a fixed format the GPU can read directly. Nvidia's RTX Neural Texture Compression (NTC), by contrast, bundles all of a material's textures into compact data from which a small network reconstructs the pixels. The NTC SDK on GitHub offers three operating modes, and this is exactly where it's decided whether memory gets saved:
| 2K material with 5 textures (per Nvidia) | Disk | Transfer (PCIe) | VRAM |
|---|---|---|---|
| uncompressed | 32 MB | 32 MB | 32 MB |
| BCn (today's standard) | 12 MB | 12 MB | 12 MB |
| NTC “on Load” | 2.5 MB | 2.5 MB | 12 MB |
| NTC “on Sample” | 2.5 MB | 2.5 MB | 2.5 MB |
With Inference on Load, textures are converted back to BCn at load time. Download size and data transfer shrink, but VRAM usage doesn't. Only Inference on Sample saves VRAM, because the network computes each texel at the exact moment it's needed. A middle ground is Inference on Feedback, which only decompresses the tiles currently visible. According to the README, this mode doesn't currently work on AMD cards.
The catch: every texture access becomes a small AI computation. For “on Sample,” Nvidia recommends at least an RTX 40-series card. According to the SDK, older GPUs manage it only “very slowly.” XDA-Developers concludes that owners of older 8 GB cards will mostly be able to use the Load mode, which saves no VRAM. Because the network delivers unfiltered pixels, “on Sample” also needs anti-aliasing or DLSS, or you'll get shimmering. Version 0.10 from RTX Kit 2026.3 comes with a blunt warning: the DirectX 12 path with matrix acceleration depends on a preview version of the Agility SDK, and for it: “DO NOT SHIP ANY PRODUCTS USING THIS.” The Vulkan paths and DirectX without this acceleration are cleared for release, but according to Nvidia they run two to four times slower without it.
Demo vs. reality: the numbers
| Case | Before | After | Savings | Cost |
|---|---|---|---|---|
| Nvidia, GTC 2026 (Tuscan villa) | 6.5 GB (BCn) | 970 MB | approx. 85% | not disclosed |
| Tom's Hardware, Sponza scene, RTX 5090, 4K | 2,041 MiB (on Load) | 303 MiB (on Sample) | approx. 85% | 1.29 → 1.38 ms (+7%) |
| same scene, RTX 5070, 1440p | 2,041 MiB | 303 MiB | approx. 85% | 1.47 → 2.00 ms (+36%) |
| Ubisoft, Assassin's Creed Mirage | – | – | approx. 30% | use limited by runtime cost |
The Tom's Hardware figures come from PC Games Hardware's summary. The baseline matters. Compared with the uncompressed 6,830 MiB, it would even be 96 percent less, which is what an earlier headline claimed. But no game ships uncompressed textures. The only honest comparison is against BCn.
The costs need a closer look. Plus 36 percent sounds dramatic, but in absolute terms it's 0.53 milliseconds. At 60 frames per second, each frame has 16.7 ms, so that's just over 3 percent of the budget. But a demo scene that finishes in 1.5 ms is doing almost nothing else. In a real game, millions of pixels per frame read multiple textures, and the cost grows with resolution and texture accesses.
That's exactly what the only real-world case shows. In February 2026, Ubisoft's research division La Forge described how Mirage ships neurally compressed textures. It saves “approximately 30 percent” of memory, and only on a selection of objects “that occur very frequently and create high memory pressure.” Ubisoft also writes: “In the shipped game, usage was limited by runtime performance.” The number of pixels the decoder is allowed to reconstruct per frame was therefore capped. Ubisoft uses its own technique with a tiny network, not Nvidia's NTC. We found no shipping game using RTX NTC. Wccftech and TweakTown don't name one either, and XDA expects the first titles no earlier than late 2026 to mid-2027. Intel has announced its own solution, Texture Set Neural Compression: up to 9x smaller with around 5 percent quality loss (measured with the FLIP image comparison metric), and up to 18x with around 7 percent. An alpha version is planned first.
In Unreal Engine 5.7, as far as we can tell, NTC is only available through Nvidia's plugin or Nvidia's own engine branch. It doesn't appear in Epic's official release notes.
What Helix and PS6 actually promise so far
At GDC in March 2026, Ronald cited a custom AMD chip for Helix, an “order of magnitude” leap in ray tracing and, according to reports, AMD's FSR Diamond with AI upscaling, multi-frame generation and neural rendering. On texture compression, he spoke of the “latest neural texture compression techniques.” At the same time, he said, they're leaning “very heavily” on Zstandard, a classic non-AI compression method for streaming from the SSD. Alpha dev kits are slated to reach studios in 2027. There's no word on memory size, price or savings figures. Windows Forum puts it well: “without needing more” shouldn't be read as “without needing anything,” and a strategy isn't a price tag.
For Sony and AMD (Project Amethyst), the three building blocks shown in October 2025 exist so far only as “promising simulations,” according to heise. The Radiance Cores are dedicated ray tracing hardware, and the Neural Arrays are compute units for AI. Only Universal Compression affects memory, and mainly the bandwidth between RAM and the GPU. According to heise, it can also save space for textures. Based on what's known so far, it isn't a neural technique.
Our verdict
In our view, neural rendering is real progress, but mainly for compute time, bandwidth and download size. Ubisoft says textures make up around 60 percent of install size, and that's where neural compression can make a big difference. As a substitute for memory chips, it has so far only worked in demos. The 85 percent comes from calm scenes on fast RTX cards. The only real-world case achieves 30 percent on a subset of objects, held back by exactly the compute time that Ronald says Helix won't need extra of.
On top of that, VRAM holds far more than textures: geometry, ray tracing structures, frame buffers and frame generation buffers, for example. Manufacturers don't publish how much of that textures account for in console games. A console maker also has to lock in memory capacity for the most demanding games of the next seven years before studios have even been able to show how widely they'll use NTC. There's no evidence that Microsoft or Sony are planning for less memory because of it. More likely, neural compression could keep demand from growing further at a time when memory is expensive.
The open question is whether the new consoles' dedicated AI hardware will make “on Sample” cheap enough for studios to use it across the board. Only when Helix or PS6 games deliver real memory figures can Ronald's claim be tested. Until then, “without additional memory” is a design goal, not a measured result.
Sources
- Pure Xbox: Project Helix information purexbox.com
- Windows Forum: Project Helix and neural rendering windowsforum.com
- Insider Gaming: Neural Rendering and Helix affordability insider-gaming.com
- Wccftech: Helix is a console that plays PC games wccftech.com
- GamingBolt: Helix leans heavily into Zstandard gamingbolt.com
- Ubisoft La Forge: Shipping Neural Texture Compression in Assassin's Creed Mirage ubisoft.com
- GitHub: NVIDIA-RTX/RTXNTC github.com
- PC Games Hardware: NTC-Test mit Tom's-Hardware-Zahlen pcgameshardware.de
- Tom's Hardware: Benchmarking RTX Neural Texture Compression tomshardware.com
- Guru3D: NTC in GTC 2026 demo guru3d.com
- XDA Developers: NTC and 8GB GPUs xda-developers.com
- Microsoft DirectX Blog: Shader Model 6.9 and the future of Cooperative Vector devblogs.microsoft.com
- PC Games Hardware: Shader Model 6.10 Preview pcgameshardware.de
- heise: AMD und Sony zeigen Technik der mutmaßlichen PlayStation 6 heise.de
- TechPowerUp: RTX Kit 2026.3 techpowerup.com
- TweakTown: Intel Texture Set Neural Compression tweaktown.com
- Push Square: Memory prices tipped to double by 2027 pushsquare.com
- Tom's Hardware: DLSS revision reduces VRAM usage tomshardware.com
- Overclock3D: Assassin's Creed neural texture compression overclock3d.net
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.