AMD FSR 5 Scarlet Cortex Preview - Published April 1st 284

AMD FSR 5 Scarlet Cortex Preview - Published April 1st

by
W1zzard
,
on ,
Manufacturer: AMD
(284 Comments) »

Introduction

AMD Logo


Spoiler: A note from the author at TechPowerUp

This article was our April Fools' Day 2026 joke. FSR 5 "Scarlet Cortex" does not exist as a product from AMD. That said, every technical concept described here is real and possible today, so if you're curious where GPU tech and DLSS/FSR/XeSS could be heading, read on. If you'd rather not "waste" your time reading about something that doesn't exist, stop now.

I've seen your feedback, and I want to address it directly. Some of you are genuinely upset, and I'm not going to brush that off.

If you feel that this article wasted your time or made you feel disrespected, I'm sorry. TechPowerUp is a trusted source you can rely on for accurate information, and I understand that publishing something fictional—even on April 1st—can feel like a violation of that trust.

Some behind-the-scenes context: in mid-March, without even thinking of April 1st, with all the news about DLSS 5, I wondered—what if AMD built something much more ambitious? Wouldn't that shift the balance of power in the GPU market? They could disrupt NVIDIA's whole business by offering better-looking games, not just more FPS, or more FPS at higher resolutions. You know me, I'm technical, so I started thinking about how this could actually be implemented from an architectural perspective. How do you feed your algorithms more/better data than the limited bits of info that upscalers currently receive from the game? I soon realized that hooking into the game's render flow is the way to go, and that the GPU driver already has the means to do exactly that. At this point I was like "OMG, this could be real"—but also realized that building such a system would be a monumental task (though entirely possible with today's technology). Around the same time, my staffers kept asking me "Wiz, are we doing an April Fools' this year?" One thing led to another, and soon I was thinking "let's do an FSR 5 review for April 1st"...

The reason I went deep on technical detail wasn't to maximize deception—it was because a shallow joke wouldn't have been interesting enough. I wanted to explore what a genuinely competitive AMD response to DLSS 5 could look like if they pushed the "adaptive learning, no game profiles" angle to its logical conclusion. Many of you said the article made you think, made you excited, made you wish this was real. That's not a failure—that was the point. And yes, when I had the first draft, it got a bit scary, even for me and my staffers, who peeked into the review and got excited that AMD Radeon was making such an incredible comeback.

To address specific concerns:

"I can't trust TPU to deliver actual news anymore." This was a once-a-year thing, published on April 1st. Every other article, every review, every benchmark—that's real, and the standards behind them haven't lapsed. We've been here for 22 years, always working hard to earn your trust, and we're not going anywhere.

"You did this for clicks and ad revenue." This article took significantly more effort than a regular review. If I wanted easy clicks, there are far lazier ways to get them. This was a creative project.

"There's no indication it's a joke." That's fair criticism. I should have added a marker sooner after April 1st had passed. I'm adding this statement now to avoid any further confusion.

"Why target AMD?" This wasn't about making AMD look bad—if anything, the article paints a very optimistic picture of what AMD could build. I chose "FSR 5" because the DLSS 5 announcement was fresh, and the technology space was ripe for speculation. Who knows, maybe we'll even see some of these ideas show up in something AMD ships down the road—and that inspiration ends up as a win for everyone.

"You fabricated images." The comparison images are based on my own real screenshots at highest detail settings (you can find the originals in the game performance reviews). I then used AI to illustrate the concept. Doing this with classic Photoshop would be impossible, and even with AI it turned out MUCH harder than I initially thought. Gemini was completely useless. Claude did better but didn't understand the direction I wanted—enhancement without changing everything around, which would break the experience with our comparison slider. ChatGPT did much better, and that's what I ended up using. I still spent more than a day refining prompts for just three screenshots. Some things, like the shop clerk in glasses, turned out worse than I wanted. Others, like the lighting and textures, look pretty impressive though, at least in my opinion.

One more thing I found fascinating about this project: I quickly realized that writing an article this detailed would itself be a monumental task, so I wondered—can AI help me? Not to write it for me, but to boost output while still producing MY thoughts and keeping 100% control over the result. I spent hours crafting the narrative. Every technical detail: driver-level texture and geometry interception, shader introspection of DXIL bytecode, INT8 inference on dedicated AI cores, mesh topology classification, adaptive online learning—these are real techniques from graphics research and driver engineering. The ideas are mine, but the AI proved to be a very clever assistant that knows a surprising amount about GPU tech (though it does hallucinate sometimes).

Making the PowerPoint slides was another challenge—I cannot believe how bad AIs are at this right now, or maybe I used the wrong tools. That alone took another day. The biggest issues were font sizes, positioning without overlaps, and creating an overall cohesive, polished look that's believable enough. I'm not 100% happy with those either, but they turned out good enough to sell the whole package.

For those wondering—we do have a long history of April Fools' jokes at TPU, though not every year. Stay tuned.

To those who enjoyed this article—thank you. To those who felt burned—I understand, and I'll keep your feedback in mind. TechPowerUp's credibility matters more to me than any single article. I will say this though: several of you told me the article taught you things about GPU architecture you didn't know before, or made you think about upscaling in a completely new way. If that happened for you, that's something you get to keep regardless of whether FSR 5 is real.

AMD's upcoming FSR 5 "Scarlet Cortex" is a real-time neural rendering technology that represents the company's most ambitious leap in visual quality to date. AMD provided us with early access to a pre-release driver build ahead of official availability, so we can bring you the full technical details and hands-on impressions. With FSR 5, AMD moves beyond traditional upscaling and frame generation into the realm of AI-driven visual enhancement—improving lighting, materials, reflections, and atmospheric effects on fully rendered frames in real time. This is AMD's direct response to NVIDIA's recently announced DLSS 5, and in several key areas, it goes further than what we've seen from NVIDIA so far.



NVIDIA's DLSS 5, unveiled at GTC in mid-March, takes a game's rendered 2D frame and motion vectors as input and uses a pre-trained generative AI model to "infuse pixels with photoreal lighting and materials." It's impressive technology, but has drawn criticism for potentially overriding artistic intent, and it requires NVIDIA's engineers to build per-game profiles—a process that doesn't scale well and may leave game developers somewhat out of the loop.

AMD's approach with Scarlet Cortex is fundamentally different. Instead of relying on a static model trained offline on curated cinematic datasets, FSR 5 uses an adaptive neural renderer that learns from the game itself as you play. There are no game-specific profiles to download and no offline training pipeline. The model ships with baseline weights trained on general rendering principles, then continuously refines its understanding of each game's visual identity—its color palette, material properties, lighting language, and artistic intent—using data gathered in real time from the rendering pipeline. Because the model adapts to whatever the game is actually doing, it attempts to inherently respect the developer's art direction rather than imposing an external notion of "photorealism."

The key capabilities of FSR 5 Scarlet Cortex are:
  • Real-time neural rendering that enhances lighting, materials, and atmospheric effects on rendered frames without altering geometry or art direction.
  • Adaptive online learning that progressively improves visual quality as you play—no game-specific profiles or downloads required.
  • INT8 quantized inference and training on RDNA 4 AI accelerators, running on dedicated hardware without competing for shader resources.
  • A deep driver-level analysis pipeline that intercepts textures, geometry, and shader programs at load time to pre-classify materials, surfaces, and rendering intent before the first frame is drawn.
  • A two-tier integration model: universal driver-level support for all games, plus an optional SDK for developers who want to provide additional rendering data to the algorithm, for significantly better results, or to limit FSR 5's abilities, for multiplayer or to preserve visual styles.
A public driver update with FSR 5 support for all RX 9000-series owners is expected later in Q2 2026. At launch, FSR 5 supports DirectX 12 titles only, but AMD tells us they are actively exploring support for Vulkan and DirectX 11 as well.

In this review, we'll explain how FSR 5 works under the hood, examine its visual impact across several games, measure the performance cost of each quality mode, and compare its output to standard rendering.

Hardware Requirements

FSR 5 Scarlet Cortex is exclusive to Radeon RX 9000-series GPUs based on the RDNA 4 architecture. The technology depends on the dedicated AI accelerators in RDNA 4, specifically their support for high-throughput INT8 matrix operations. The neural rendering model runs entirely on these AI cores, separate from the shader units, so it does not directly compete with the game's rendering workload.

The reason older RDNA generations cannot support FSR 5 is not simply a matter of inference speed. The adaptive learning system performs both neural inference and weight updates—effectively running training and inference simultaneously at 60+ FPS. This dual workload requires hardware-accelerated INT8 matrix throughput that only RDNA 4's AI accelerators can deliver. There is no shader-based fallback; the computational demands are too high for software emulation at real-time frame rates.

FSR 5 sits alongside the existing FSR 4 feature set (ML upscaling, frame generation, ray regeneration, radiance caching) and can be used in combination with any of them. It operates as a post-process enhancement stage that runs after the game's rendering pipeline but before final output, making it compatible with upscaling and frame generation workflows.

At launch, FSR 5 uses a whitelist—FSR 5 Neural Rendering only activates for games AMD has explicitly validated and added to their driver profile database. The supported game list will grow with each driver update. AMD has told us they plan to eventually transition to a blacklist approach once the technology matures, where FSR 5 is enabled by default in all games except known competitive multiplayer titles that AMD maintains on an exclusion list. The feature is controlled through the Radeon Software overlay, where users can select from three quality/performance presets and monitor the real-time status of the adaptive learning system.

How It Works — Driver-Level Pipeline


One of the most impressive aspects of FSR 5 is how much the driver can figure out about a game's rendering intent without any developer cooperation. AMD's driver already sits in the middle of every draw call, every texture upload, and every shader compilation—Scarlet Cortex simply puts that position to use.

Our Patreon Silver Supporters can read articles in single-page format.
Return to classic article view
Nov 15th, 2026 00:37 CST change timezone

Latest GPU Drivers

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts