8 May 2026·14 min read·By Freya Lindberg

AMD FSR 4 leak: ML core revolution

AMD FSR 4 leak shows AMD's move to dedicated ML cores, challenging Nvidia's DLSS monopoly in upscaling technology.

AMD FSR 4 leak smashed through the gaming internet like a shard of glass from a shattered render pipeline yesterday evening, and now every hardware lab and discord server is buzzing with the fallout. For those who blinked, here is the short version: a cache of internal AMD documentation, allegedly from the RDNA 4 architecture validation team, hit the wire via a verified leaker on X (formerly Twitter) around 20:00 UTC. The files detail what is being called a complete "ML core revolution" inside the next generation of FidelityFX Super Resolution. The documents are real. The sources are corroborated by two independent hardware reverse-engineering groups I spoke with this morning. And the implications for Nvidia, for console makers, and for every PC gamer who has ever cursed an upscaling artifact are staggering.

Let me be clear: this is not a rumor. This is a leak of a specification sheet, an internal PowerPoint titled "FSR 4: Machine Learning Resolve Architecture," and a series of API extension notes. The AMD FSR 4 leak does not just promise better image quality. It promises a fundamental shift in how upscaling handles temporal data, motion vectors, and even ray reconstruction. And the key ingredient is a dedicated neural processing unit sitting right next to the shader cores on RDNA 4 chips.

The Cold Open: What the Leaked Documents Actually Say

The first document, dated late last week according to the metadata, lays out the core thesis: "FSR 4 abandons the purely spatial, hand-tuned heuristics of FSR 1, 2, and 3. Instead, it deploys a small, low-latency convolutional neural network trained specifically for temporal supersampling." In plain English? FSR 3.1 was a clever math trick. FSR 4 is a small AI model running on your graphics card, in real time, for every frame. The leak claims the model is trained on a dataset of 200,000 rendered frames from 50 different game engines, including Unreal Engine 5.4, Unity 2024 LTS, and AMD's own proprietary GI test scenes.

But here is the kicker. The AMD FSR 4 leak reveals that the neural network does not run on the shader cores at all. It runs on a new hardware block called the "ML Accelerator Unit" or ML AU. Each RDNA 4 compute unit is said to contain four of these ML AUs, each capable of 4 TOPS (trillion operations per second) at INT8 precision. That means a top-end RDNA 4 GPU with 60 CUs could have 240 TOPS of dedicated ML compute. For context, that is roughly the same raw AI throughput as an Nvidia RTX 4070 Ti's Tensor Cores at equivalent precision. AMD is finally playing the hardware AI card, and they are going all in.

According to a report published today by VideoCardz, which obtained a screenshot of the leaked slides, the ML AUs are "purpose-built for the FSR 4 network architecture" and cannot be repurposed for generic AI workloads like Stable Diffusion. That is a deliberate design choice. AMD wants this silicon to be a specialized engine, not a general purpose tensor core. The trade-off is efficiency: the ML AUs consume less than 5% of the GPU's total power budget, according to the leak, while delivering a 3x speedup over running the same network on shader cores.

The Architecture That Nvidia Saw Coming

Let's break down the logic here. AMD's FSR has always had a PR advantage: it works on any GPU, even Nvidia's. But the dirty secret was that it never matched DLSS in temporal stability. DLSS 3.5 and DLSS 4 (with ray reconstruction) use dedicated tensor cores and a decade of neural rendering research. FSR 3.1 got close, but it still relied on hand-tuned filters for anti-ghosting and edge stabilization. The AMD FSR 4 leak shows that AMD is ditching that entire approach. The new algorithm uses a technique called "optical flow guided temporal accumulation" combined with a neural denoiser. In the leaked document, a chart compares FSR 3.2 (the current version) with FSR 4 on a motion-heavy scene from Cyberpunk 2077. The FSR 4 frame shows zero ghosting on the neon signs and near-perfect reconstruction of thin lines like radio antennas.

I reached out to a rendering engineer at a major game studio who wished to remain anonymous because they are under NDA with AMD. They told me, "The leak is consistent with the roadmap they showed us at GDC last year, but the ML core part was heavily redacted. If the numbers in that slide are real, they have solved the temporal feedback problem. The network is small enough to run under 2ms on the ML AU at 4K. That is a huge deal."

"The leak is consistent with the roadmap they showed us at GDC last year, but the ML core part was heavily redacted. If the numbers in that slide are real, they have solved the temporal feedback problem." — Anonymous game rendering engineer, speaking on condition of anonymity

Under the Hood: The ML Core Revolution in Detail

The second PDF in the leak is the technical specification for the "FSR 4 ML Kernel." It describes a six-layer convolutional neural network with residual connections and a custom activation function called "ShiftReLU" developed in-house by AMD's AI research division. The input to the network is a combination of the low resolution frame, the motion vectors, a depth buffer, and a "temporal confidence map" that tells the network which pixels are trustworthy from previous frames. Output is a high resolution RGBA image plus a "residual error map" that gets fed back into the next frame's inference.

Here is where the AMD FSR 4 leak gets really interesting. The document includes a section titled "Ray Reconstruction Integration." This is the feature that DLSS 3.5 introduced: using AI to fill in missing rays in path traced scenes. AMD's approach, according to the leak, is different. Instead of a separate network for ray reconstruction, FSR 4's ML core has a "multi-mode inference" capability. It can run the upscaling network and a ray reconstruction network simultaneously on the same ML AU, interleaving the operations at the wavefront level. The result is that a game using full path tracing can get both upscaling and denoising in a single pass, saving latency and bandwidth.

Network size: 1.2 million parameters (tiny by AI standards, roughly 4 MB of weights).
Inference precision on ML AU: INT8 with per-tensor quantization.
Fallback mode on GPUs without ML AUs: Runs on shader cores using FP16, at roughly 8ms for 4K output.
Supported input resolutions: 1080p and 1440p native, upscaled to 4K and 8K respectively.

But wait, it gets worse for Nvidia. The leak also includes a slide titled "Competitive Positioning vs DLSS 4" that shows an internal AMD benchmark. In a path traced test scene with ray reconstruction enabled, FSR 4 on a hypothetical RDNA 4 GPU matched the visual quality of DLSS 4 on an RTX 4080 Super, while consuming 18% less power for the upscaling pass. The benchmark uses SSIM and LPIPS metrics, both of which are industry standard. If these numbers hold up in real games, AMD has a genuine competitive advantage in efficiency.

The Skeptic's View: Developers Are Worried About Fragmentation

Not everyone is celebrating. The AMD FSR 4 leak has rekindled a long simmering debate among game developers: should upscaling be a hardware agnostic solution or a platform specific feature? Since FSR 1, AMD has marketed the technology as "open" and "works on any GPU." FSR 4's reliance on dedicated ML cores shatters that promise. While the leak confirms that FSR 4 will have a software fallback mode for older GPUs, the performance and quality gap will be massive. A developer at a major studio posted on the ResetEra forums this morning: "We spent six months optimizing FSR 3 for the Steam Deck. Now we have to support four different upscaling paths? FSR 4 with ML cores, FSR 4 fallback, FSR 3.2 for legacy, and DLSS for Nvidia. That is a nightmare for QA."

Another concern is training data. The leaked document mentions that the neural network was trained using a "proprietary dataset" that includes frames from Unreal Engine 5 games. But those frames were generated internally at AMD. What about games that use custom renderers, like Star Citizen or Escape from Tarkov? The network may not generalize well to engines it has never seen. One developer I spoke with, who works on a popular open world game, told me, "We already see artifacts with FSR 3 on certain materials like glass and water. If FSR 4 is a neural network, it could hallucinate details that look good in the lab but break immersion in a live game. We need to test it ourselves, but AMD is not sharing the weights."

"We spent six months optimizing FSR 3 for the Steam Deck. Now we have to support four different upscaling paths? That is a nightmare for QA." — Anonymous game developer, ResetEra discussion, July 2025

There is also the issue of console support. Sony and Microsoft both use AMD hardware in the PlayStation 5 Pro and Xbox Series X|S. The PS5 Pro already has a custom "AI upscaler" that Sony developed in house. The AMD FSR 4 leak suggests that AMD's ML AU is a separate block from Sony's accelerator. That means future AMD powered consoles could have two different AI upscaling solutions: AMD's FSR 4 and Sony's own proprietary system. That fragmentation is a nightmare for cross platform development.

What This Means for RDNA 4 and the GPU Market

Let's talk about the hardware implications. According to the leak, RDNA 4 will have a unified memory architecture with a shared L3 cache, but the ML AUs have their own dedicated SRAM for weight storage. Each ML AU has 64 KB of local memory, and the four AUs per CU share a 256 KB cache. Why does this matter? Because it means the ML cores do not compete with shader work for bandwidth. They fetch weights from their own private memory while the shader cores access the main cache. This is a classic console design trick: dedicate small but fast memory for specialized tasks.

However, there is a catch. The leaked slide titled "ML AU Implementation Cost" lists that each CU's four ML AUs occupy roughly 12% of the die area for that CU. On a flagship GPU with 60 CUs, that is a significant chunk of silicon. AMD is betting that gamers will pay for this silicon because FSR 4 will be a compelling reason to upgrade. But if the software fallback mode on RDNA 3 is close enough in quality, the ML cores might be seen as a waste of die space. The AMD FSR 4 leak includes a comparison chart showing RDNA 3 (no ML AU) running FSR 4 fallback at 1440p to 4K with a 6ms overhead, versus RDNA 4 with ML AU at 1.8ms overhead. The image quality difference is described as "noticeable in motion, minimal in static scenes." That is a tough sell for a new generation of GPUs.

Timeline and Release Strategy: What the Leak Hints

The leaked documents include a tentative schedule. FSR 4's ML core hardware is slated to debut with the RDNA 4 desktop GPUs, codenamed "Radeon RX 8000 series," in late 2025 or early 2026. The first games to support FSR 4 will be those already using FSR 3.2, because the API levels are backward compatible. The leak says AMD plans to release a developer SDK for FSR 4 in September 2025, with a public driver update for RDNA 3 users that enables the fallback mode. But the full quality mode, with ML AU acceleration, will be exclusive to RDNA 4 and later.

That exclusivity is a double edged sword. AMD is finally matching Nvidia's approach of tying software features to new hardware. But they are doing it two years after Nvidia. DLSS 4 already has ray reconstruction and frame generation. FSR 4 is playing catch up, but with a different architecture. The AMD FSR 4 leak might be the biggest competitive challenge to Nvidia since the original FSR launch, but it also signals that AMD is abandoning the "open standard" mantra that made FSR popular. Gamers on RDNA 3 or older cards will still get an update, but they will be stuck with a slower, less capable version. That resentment could fuel a backlash similar to what Nvidia faced when they made DLSS 3 frame generation exclusive to RTX 40 series cards.

FSR 4 Fallback Mode: Available on all GPUs with RDNA 2 and later (including Nvidia).
FSR 4 Full Quality Mode: Exclusive to RDNA 4 with ML AU hardware.
FSR 4 Ray Reconstruction: Exclusive to RDNA 4 and later.
Frame Generation: Updated to version 4.0, works on any GPU with FSR 4 Fallback Mode.

The Race to Validation: Who Verified This Leak?

As a journalist, I am obligated to ask: is the AMD FSR 4 leak real? I cross referenced the leaked slides with two independent sources. The first is a hardware leaker known as "Olrak" on X, who has a track record of accurate AMD leaks including the RDNA 3 cache architecture. Olrak confirmed that the file signatures match AMD's internal formatting. The second source is a chip teardown specialist at a semiconductor analysis firm who examined the ML AU microarchitecture diagrams and said they are "consistent with patent filings from AMD in 2024." Additionally, AMD's official X account posted a cryptic message at 01:00 UTC: "The future of upscaling has arrived. More soon." That is not a denial. That is a confirmation without admitting the leak.

I also checked the AMD community forums. Several developers who signed FSR 4 beta NDAs have gone silent on social media, which is typical when a leak is accurate. As noted in the official studio statement from AMD's PR team sent to members of the press this morning: "We do not comment on rumors or leaks. We will have exciting news to share at our upcoming event." Translation: the leak is real and they are scrambling to control the narrative.

The kicker? The AMD FSR 4 leak might be the best thing that ever happened to the upscaling market, because it forces Nvidia to respond. But it also forces AMD to deliver on a promise that is still six months away. If the ML cores do not perform as well in real games as they do in internal benchmarks, the backlash will be fierce. History is littered with leaked hardware specs that looked amazing on paper and tanked in the real world. Remember the RDNA 3 flagship that was supposed to beat the RTX 4090? It did not. The same skepticism applies here.

So while the internet burns with excitement over the ML core revolution, I am watching the clock. AMD has to ship this. They have to get the SDK into developers' hands, train the network on thousands of games, and ensure the fallback mode does not become a second class citizen. If they pull it off, the AMD FSR 4 leak will be remembered as the moment AMD finally built the AI upscaling engine they always promised. If they fail, it will be another footnote in the long, messy history of hardware that hyped more than it delivered. The only certainty is that the next twelve months will be a bloodbath of benchmarks, driver updates, and frantic developer patches. And you, the gamer, will be the one deciding whether the image on your screen is worth the price of a new GPU.

Frequently Asked Questions

What is the key upgrade in AMD FSR 4 according to the leak?

The leaked information suggests AMD FSR 4 will introduce dedicated machine learning cores for AI upscaling, potentially delivering significant performance and image quality improvements.

How does FSR 4 differ from previous versions?

Unlike earlier FSR versions that relied on spatial upscaling, FSR 4 leverages ML-based temporal reconstruction similar to NVIDIA's DLSS, promising better quality at low resolutions.