AMD will finally join Nvidia in FSR 4 with AI powering frame generation. However, there is no word yet on upscaling.

FidelityFX super resolution, AMD's answer Nvidia DLSS has been around for a bit over three years and it's a real boon for all kinds of GPUs including Nvidia GeForce models. The latest version of FSR 3.1 offers a shader based upscaler system and frame generation system. However, with the tech world enthralled by AI, AMD decided that it couldn't ignore this anymore and announced that its next iteration would be AI-based.

In an interview with Tom's Hardware Jack Huynh, AMD's senior Vice President, discussed some of the frustrations associated with playing on handheld gaming computers. He also mentioned that a high framerate greatly limited your ability to play.

"My number one priority on the handheld side is battery life. Battery life is a problem with the Lenovo Legion Go or the Asus ROG Ally. I need multiple hours. "I need to play [Black Myth] Wukong over three hours and not 60 minutes."

He's not mistaken. Even with the ROG ally X's huge battery, I rarely get to play for more than an hour in performance mode. While you don't have to run the chip at full speed in every game, it is necessary for any new games with high-end graphics. Huynh believes that the solution will come in the form the next version FideltyFX Super Resolution.

"This is the FSR 4 we're adding." We completely restructured the team around 9-12 months ago, to AI-based. Now we're using AI-based frame interpolation and frame generation to maximize battery life. We could then lock the frames per seconds, maybe it's 35 frames per second or 30."

Huynh's only comment was that it's all AI-based. We're left to speculate and ponder. Let's begin with the main reason for making FSR4 AI-based, which is to improve battery life. It's admirable, but I hope that AI will also help to solve FSR 4's visual shortcomings (watch the video below).

Unlike Intel Arc or Nvidia RTX, there is no specific hardware in the GPU of a handheld gaming PC, Radeon-equipped laptop, that can accelerate neural network calculations. The Ryzen AI 300 Series does have a NPU (neural processor unit) that can handle such workloads while using relatively less power.

Does that mean FSR 4, like the current version, will require access to a matrix/tensor core or will it be entirely shader-based? It's not necessary to have matrix/tensor processors in order to run AI algorithms. However, offloading these tasks onto them allows the standard ALUs to be used for shader routines.

Does this mean that FSR 4 is not accessible to anyone using a different system if FSR 4 uses an NPU, or AMD's new generation of GPUs based on the RDNA 4 architecture has a small amount of matrix units? FSR's best feature is that it can run on any GPU: it only needs to support a minimum level of compute shaders in Direct3D and Vulkan.

I'd be surprised if AMD planned to abandon this approach, which leads me suspect that the AI nature FSR 4 will not be as reliant upon hardware support as XeSS or DLSS is. Intel's upscaler has two modes, one for all GPUs including FSR, and another for Arc GPUs that uses the matrix units.

Here's what I believe: FSR 4 is still shader-based, but will use an AI-based frame generation system that can potentially leverage different hardware (GPU, CPU, NPU).

It could be similar to its RDNA 2 or later GPU-based Fluid Motion Frames System. It could be AFMF, which then uses a neuronal network to clean up an interpolated frame.

AMD has said it will eventually roll out AI upscaling on all of its gaming devices, so maybe every aspect of FSR 4 is AI-powered. Your guess is as good mine.

We won't know much more about FSR 4 until AMD is ready. Let's have some fun speculating what it might be until then. Anyone want to place a wager on it being a mixture of ChatGPT with Stable Diffusion. "Create a picture like these two frames, but with one frame in the past."