AMD plans to extend FSR 4, its AI-based technology for enhancing game visuals, to APUs and handheld gaming PCs at the end of 2026. According to the Korean outlet The Elec, Jack Huynh, who leads AMD's computing and graphics business, said so at a press briefing in Seoul on October 7. He said AMD has developed a new, dedicated lightweight model for existing APUs.
AI upscaling, which has so far been ahead on desktop GPUs, may now reach handhelds that people already own. But even AI that reconstructs sharp images takes computing time. On handhelds, which must run games on limited power, how light that processing can be made while keeping image quality will shape how usable the feature is.
Two Paths: Existing APUs and New Products
Huynh named APUs, gaming laptops and handheld devices as year-end targets, and explained that AMD will proceed in two directions: support for existing APUs and the launch of new product lines. An APU is a processor that puts a CPU and GPU on a single chip, and it handles graphics in handheld gaming PCs. This statement is therefore not a plan limited to new devices.
Existing APUs will be supported by a dedicated model slimmed down for APU environments. The development goals Huynh cited were excellent image quality and low latency. However, the method of slimming down, the APU model numbers that will be supported and the release date have not yet been disclosed.
On June 24, 2026, AMD had already expanded support for FSR Upscaling 4.1.1 to discrete GPUs in the RDNA 3-generation Radeon RX 7000 series through FSR SDK 2.3. For RDNA 3.5 and others, it provides a switch to the conventional approach that does not use machine learning. Supporting the same FSR SDK does not necessarily mean a chip is actually reconstructing images with an AI model.
FSR 4 is now officially called "FSR Upscaling" and, together with frame generation and other features, makes up FSR "Redstone." According to AMD's current guidance, machine-learning upscaling targets the RX 7000 and RX 9000 series, with the RX 6000 series planned for 2027. The year-end remark about expanding to more products should not be read as a promise that all past generations will be supported at once.
Why a Lightweight Model: The Per-Frame Time Budget
In a game running at 60fps, the time available per frame is about 16.7 milliseconds. Upscaling renders the game at a lower resolution and reconstructs the higher-resolution image shown on screen. FSR 4 uses machine learning to process information from multiple frames and motion data passed from the game, filling in fine detail.
Part of the time saved by rendering at a lower resolution is spent on reconstructing the image. A clue to how that balance works out is the processing times in AMD's developer documentation. Converting the published reference values for 1080p output into a share of a 60fps frame gives the following.
| GPU | FSR 4 processing time | Share of a 60fps frame |
|---|---|---|
| Radeon RX 7600 | 1.998 ms | About 12% |
| Radeon RX 9070 XT | 0.352 ms | About 2% |
In AMD's reference values, FSR 4 processing alone at 1080p output uses about 12% of a 60fps frame on the RX 7600 and about 2% on the RX 9070 XT.
The table uses values for FSR Upscaling 4.1.1 in Performance mode, checked on October 9, 2026, with RCAS sharpening disabled and the GPU running at peak clock. The percentages were calculated as processing time ÷ (1000 ÷ 60) × 100. This is the total reconstruction time; it is neither the added time compared with FSR 3 nor the delay between input and on-screen response. The GPU configurations also differ, so the gap cannot be treated as purely a generational difference.
The performance of a handheld lightweight model cannot be predicted from this table. Still, it shows that using AI image enhancement at low power requires keeping the time spent on reconstruction down. AMD's documentation also explains that FSR processing time depends mainly on output resolution. Lowering the internal rendering resolution does not reduce the reconstruction load by the same proportion.
For example, at 1080p output, Quality mode, which scales each dimension by 1.5, renders internally at 720p, with about 44% of the pixels of the display resolution. Performance mode, at 2x per dimension, renders at 540p, or 25% of the pixels. These are pixel ratios calculated from the scaling factors, not a claim that the game's overall load falls by the same proportion. The less that is rendered, the more the quality of reconstruction that preserves detail from less information matters.
Image Quality Gains vs. Differences in fps and Responsiveness
What can be expected from moving to FSR 4 is, first of all, better images built from a low internal resolution. Compared with rendering every pixel at native resolution, lowering the resolution saves rendering time. But for games that already use FSR 3.1 at the same internal resolution, the speed after switching to FSR 4 also depends on the reconstruction load. Better image quality and higher fps are separate outcomes.
The value of a lightweight model lies in whether it can create a combination that keeps image quality convincing even at lower internal resolutions, with the rendering savings outweighing the reconstruction load. If achieved, one could use it to raise fps at the same power, or to cap fps and cut power. How much the latter extends battery life, though, must await measurements with image quality and fps caps matched.
Frame generation also plays a different role from upscaling. It interpolates images between existing frames to make the display smoother; even if more interpolated images appear, the game is not processing new input at that rate. AMD's explanation of FSR 3 likewise states that interpolation adds latency and that mechanisms are built in to reduce it.
FSR 3's developer guidance also recommends a pre-interpolation frame rate of at least 60fps, because visual artifacts from interpolation are more noticeable at low fps. This is guidance for the existing FSR 3, not a requirement for the unannounced lightweight FSR 4, but it shows why the displayed fps number alone cannot determine how comfortable a handheld feels.
Intel's 42% and "Over 2x": Different Comparison Conditions
Intel announced its Arc G3 series for handhelds on May 28 and is promoting XeSS 3 as a combination of AI super resolution, multi-frame generation and low-latency technology. Even if AMD introduces a lightweight FSR 4, competition will not be decided by whether an upscaler exists alone.
Intel's claim of being "42% faster than the Ryzen Z2 Extreme" needs to be distinguished, based on the company's test conditions, from the separate claim of "over 2x smoother."
| Intel's claim | Number of titles | Image processing conditions |
|---|---|---|
| Arc G3 Extreme is 42% faster | 36 titles | 2x upscaling used in supported titles |
| Arc G3 Extreme is over 2x smoother | 15 titles | Super resolution plus XeSS/FSR frame generation or XeSS multi-frame generation where supported |
Both are Intel measurements as of May 28, 2026, with sustained power of 35W on both devices, games at 1080p and High settings, and average performance ratios summarized with a geometric mean. The comparison was between a prototype MSI Claw 8 EX AI+ and the ASUS ROG Xbox Ally X. Memory configurations also differ: 32GB of LPDDR5X-8533 for the former versus 24GB of LPDDR5X-8000 for the latter. The table does not show absolute fps for each title.
The 42% row does not state that frame generation was enabled; it is a comparison using upscaling. It cannot be called a difference in native-resolution rendering performance, nor explained by frame generation alone. Moreover, even if setting names and display resolutions are the same, the output image quality of different super-resolution technologies is not necessarily the same.
How far a lightweight FSR 4 narrows this gap will not be known until the same games are measured with image quality and power matched. AI super resolution reaching existing AMD devices would widen the options for image quality, but its ability to counter Intel will be judged by the rendering performance and responsiveness that remain when that quality is kept.
Game and Device Distribution Conditions Remain Open
Current FSR Upscaling has two routes: games that integrate it directly, and games supporting FSR 3.1 or later that are updated through AMD Software. According to AMD's guidance, the latter replaces the processing DLL to switch to the machine-learning version. Even if the hardware is supported, not every game will automatically be.
Whether the year-end lightweight model for APUs will be delivered through the same route has not been announced. Which ROG Ally and Legion Go models will be supported, and whether it will work on Steam Deck or SteamOS, were also not settled in these remarks. The specific GPU generation of the new APUs is not clear either.
Support also needs to be separated by feature. What expanded to the RX 7000 in the current SDK is machine-learning upscaling; machine-learning frame generation and ray regeneration are for the RX 9000 and later. FSR 4 becoming available on APUs does not mean those features will arrive with it.
When it is offered at year-end, it will be worth looking at not only the names of supported devices but also image quality in motion and pre-interpolation fps. Measuring the delay between operating the controls and seeing the display change, and comparing power consumption with fps caps matched, would show what the lightweight model actually improved on handhelds. If image quality and lightness of processing prove compatible, it could widen the settings available to play on handhelds people already own, without having to buy a new one.