Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

Apex Compute Releases DLSS Alternative for GTX 16 and Radeon: Frame in 1.01 ms vs 1.26 ms

Apex Compute has released a beta of ANR FG, a frame generator that uses a neural network to create intermediate frames between real ones for smoother gameplay. It replaces NVIDIA's DLSS Frame Generation on GTX 16, Radeon, Arc and Apple Silicon. According to the company's internal measurements, one generated frame at 4K on the RTX 5090 takes 1.01 ms, compared with 1.26 ms for DLSS. There are no independent tests yet, so these figures remain the manufacturer's claims.

AI-processed from Habr AI; edited by Hamidun News
Apex Compute Releases DLSS Alternative for GTX 16 and Radeon: Frame in 1.01 ms vs 1.26 ms
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The neural network weighs 0.17 MB and needs 214 MB of VRAM, versus 562 MB for DLSS.
It installs through OptiScalercopy the files into the game folder and enable DLSS FG.
Key metric #3The source code on GitHub is still closed, with publication promised later.

If a game runs at 40 real frames per second, a neural network can generate another 40 intermediate frames between them, making the picture noticeably smoother. This is how DLSS Frame Generation from NVIDIA works, but it is not available to everyone: owners of GTX 16, older Radeon cards or Macs either lack the feature or get worse performance from it. Apex Compute has released an open beta of its alternative, Apex Neural Rendering Frame Generation, or ANR FG.

It is injected into games in place of DLSS and runs on graphics cards from different manufacturers. The headline figure from the company's announcement is the time needed to produce one generated frame. On a GeForce RTX 5090 at 4K (3840 × 2160), ANR FG takes 1.

01 milliseconds per frame, while the standard DLSS Frame Generation takes 1.26 milliseconds. The lower this time, the less the graphics card is distracted from the game's main rendering.

How it works and what it runs on

ANR is designed as a set of three modules: resolution upscaling, ray tracing reconstruction and frame generation. So far only the last one is available. The upscaling and ray denoising models are still being trained.

ANR FG does not enter games as a native plugin but through the OptiScaler layer. It intercepts calls to the standard DLSS Frame Generation and launches the Apex neural network instead. For users the setup is simple: copy the files into the game folder, install the Microsoft Visual C++ 2015–2022 Redistributable (x64), and enable NVIDIA's regular frame generation in the settings. The game believes it is running DLSS, while the frames are actually drawn by another model. There are two build options: a full package with OptiScaler included, and a lighter add-on for those who already have the layer installed.

The list of supported hardware is broad:

  • GeForce RTX 20, 30, 40 and 50 series, as well as GTX 16 without tensor cores;
  • Radeon RX 5000, 6000, 7000 and 9000 series, plus Steam Deck;
  • Intel Arc A and B series;
  • Apple Silicon from M1 onward on macOS 26 and above, via Metal.

On PC, Windows 10 and 11, Linux and SteamOS via Proton are listed. For Macs there is a separate test bench where ANR FG can be compared with the built-in MetalFX technology.

What the numbers claim

Apex puts its main emphasis on compactness: the neural network takes up just 0.17 MB. ANR FG requires 214 MB of VRAM versus 562 MB for NVIDIA's implementation, roughly 62% less. For portable devices with limited memory, this difference may matter more than fractions of a millisecond.

In its summary tables the company claims an advantage of up to 2.4 times in the speed of the generation stage. On the Radeon RX 7800 XT, Intel Arc B580 and Arc A770, ANR FG is faster than AMD FSR 3.1. On the Radeon RX 9060 XT the result is mixed. In 2X mode (one generated frame for each real one) it is faster than FSR 3.1, while at three- and four-times generation FSR 4 with machine learning comes out ahead.

Apex shows quality on frames from The Last of Us Part II Remastered and Hogwarts Legacy, using three video quality metrics: CGVQM-5, ColorVideoVDP and PSNR. In 2X mode ANR FG scores 96.59 against 93.66 for DLSS on CGVQM-5, and 33.84 dB against 31.22 dB on PSNR. The higher the value, the closer the generated frame is to the real one. In every table the company has published, its solution comes out ahead.

Caveats worth keeping in mind

First, all measurements were performed by Apex Compute itself. There are no independent comparisons yet, and in a manufacturer's tables the win almost always goes to its own product.

Second, the "open" beta is open only in name for now. The GitHub repository contains documentation and compiled files, but the company has promised to publish the neural network's source code later. The community cannot currently inspect the model or improve it.

Third, substituting libraries through OptiScaler always carries risks. It is unclear how ANR FG will behave in games where performance is limited by the CPU, or whether artifacts will appear in fast-moving scenes: blurring at object edges, interface jitter, input lag. These are common problems for any frame generation, and independent testers need to check them separately.

What to do in practice

If you have a GTX 16, an older Radeon or a Mac and want to try DLSS-style frame generation, the beta is a cheap experiment: the files are copied into the game folder and just as easily removed. Developers and those tracking the market should take a close look at the approach: a tiny neural network layered over an interception layer shows that NVIDIA's technologies are getting cross-platform competitors. For now, wait for third-party measurements and the code release, and treat Apex's numbers as the manufacturer's claims.

Story

Key Questions Answered

How it works and what it runs on?

ANR is designed as a set of three modules: resolution upscaling, ray tracing reconstruction and frame generation. So far only the last one is available. The upscaling and ray denoising models are still being trained. ANR FG does not enter games as a native plugin but through the OptiScaler layer. It intercepts calls to the standard DLSS Frame Generation and launches the Apex neural network instead.

What the numbers claim?

Apex puts its main emphasis on compactness: the neural network takes up just 0.17 MB. ANR FG requires 214 MB of VRAM versus 562 MB for NVIDIA's implementation, roughly 62% less. For portable devices with limited memory, this difference may matter more than fractions of a millisecond. In its summary tables the company claims an advantage of up to 2.4 times in the speed of the generation stage. On the Radeon RX 7800 XT, Intel Arc B580 and Arc A770, ANR FG is faster than AMD FSR 3.1.

What are the key details on Caveats worth keeping in mind?

First, all measurements were performed by Apex Compute itself. There are no independent comparisons yet, and in a manufacturer's tables the win almost always goes to its own product. Second, the "open" beta is open only in name for now. The GitHub repository contains documentation and compiled files, but the company has promised to publish the neural network's source code later. The community cannot currently inspect the model or improve it. Third, substituting libraries through OptiScaler always carries risks.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…