Today ARM is announcing it’s Neural Unit and various GPU upscaling technologies that use it. SemiAccurate welcomes ARM to the upscaling club, the more the merrier.
We will kick this one off by saying that at times, scheduling sucks. ARM is announcing their Neural Unit (NU) and three things that use it, Neural Super Sampling (NSS), Neural Frame-Rate Upscaling (NFRU), and Neural Super Sampling Denoising (NSSD). The first two are being talked about today, the latter two are coming later. What does this have to do with scheduling? The normal ARM yearly core/device reveal hasn’t happened yet so these technologies are going to be in a chip next year that they can’t talk about. Unfortunately they can’t give out details of the hardware either so we can’t tell you much about how it works, just what it does. That said on with the show.
Today we have Neural Super Sampling which runs on the Neural Unit in the GPU. Like the variants of the technology from Nvidia and AMD, it takes a lower rez input and uses ‘AI’ to upscale it hopefully cleanly. The end result is similar quality to a higher rez render at a fraction of the compute, and therefore energy and time cost. Some call this fake frames and, well, we agree but the tech seems here to stay. If used properly it can make things look better, especially on underpowered hardware, but opens a can of worms about what the performance of the hardware actually is. Feel free to debate this on your own, that is what the net was made for, right?
Imagine ASR where NSS is
You might recall the slide above from earlier in the year when the box that said ARM NSS said ARM ASR or Accuracy Super Resolution. NSS does the same thing with more features and better because AI. AI may or may not be the cause but NSS does seem better than ASR and is possibly more efficient, more on that when ARM reveals the hardware. In any case it does clearly show that dedicated hardware can do things better than generic shaders. It will be interesting to see the nuts and bolts of how ARM implemented the tech, hopefully it won’t be long.
ARM NSS in the Vulkan pipeline
The NU is part of the GPU and that allows ARM to tie NSS into the graphics pipeline with standard Vulkan extensions. It adds another pipeline to the normal graphics and compute pipelines which do the AI upscaling, then feeds things back to the graphics pipeline. How exactly this is done determines how well the whole scheme works and that hasn’t been revealed. Yet. Soon. We hope.
ARM NSS model architecture
As you can see above, the whole process works similarly to most AI networks, layers of compute that feed to each other. Since the NU is part of the GPU itself, this should be fairly seamless and transparent, ARM claims ~4ms per frame for the process so that is well under the time limit for even 120FPS devices. If the claims about NSS are true, the tech should also speed up rendering per frame, lower internal bandwidth, use less memory, and make birds sing happier tunes. Again more when they reveal the hardware but the claims are realistic so far. One interesting bit they did reveal was that NSS is aware of GPU load and can scale it’s resource usage using it as an input.
The best part of all of this is ARM is fully opening NSS. About the time you read this, ARM will have presented their tech at Siggraph 2025 and put out a full SDK and PC emulator. If your GPU runs Vulkan, you can play with NSS too. If you want to play with NSS on native hardware, you need an ARM GPU with an NU which is currently listed as ‘2026 GPU’. If you want the SDK and emulator, it should be up on the ARM web site by now.
At the beginning of this tale we mentioned three technologies but haven’t really talked about NFRU or NSSD yet. This is because they aren’t being revealed today but should be in the not so distant future. NFRU basically interpolates new frames between rendered ones and NSSD denoises images, if you are familiar with similar technologies from Nvidia and AMD, you know the drill. Since these two use the same hardware as NSS, we expect them to run on the mythical 2026 GPU as well. SDKs for both should be out long before the hardware itself, and in keeping with ARM tradition, both should be open as well. Good job keeping things open people.
So in the end we have a new Neural Unit that does ‘AI’ upscaling better than the old way. NSS has an open Vulkan API that should be out now plus a full PC emulator to play with until hardware arrives. Two related technologies will be released in coming months and best of all, the hardware reveal will tell us how this is all tied together. For once we can’t wait to hear the details on something AI related.S|A
