Tom Petersen, an Intel fellow on graphics architecture, sat down with Gizmodo in downtown Taipei last week and described a technology Intel has been building since 2023: frame extrapolation, an AI system that predicts where a player will move and synthesizes a display frame before the GPU has finished rendering the real one. The pitch is smoother gameplay without the input-lag penalty that has followed every AI frame-generation system since NVIDIA introduced the interpolation method in 2022.
Every AI frame-generation method in wide use today adds latency by design. The interpolation model underpinning DLSS 3 (NVIDIA’s Deep Learning Super Sampling frame-generation feature), AMD’s FSR 3 (FidelityFX Super Resolution 3), and Intel’s own XeSS MFG (XeSS Multi-Frame Generation, which boosts apparent frames-per-second counts by inserting AI-synthesized frames) requires a second rendered frame before it can generate anything. Extrapolation would skip that wait. Petersen told Gizmodo the technology is “almost” ready for a public showcase, but it wasn’t finished in time for Computex 2026.
Why Frame Gen Keeps Players Waiting
Standard frame interpolation requires two genuine rendered frames. The algorithm analyses motion vectors between them and inserts a synthesized image into the display interval. Output frame rate goes up, at least on paper.
Frame one gets rendered and then sits idle while frame two is produced. The player’s input at the moment of frame one isn’t reflected on screen until the pair is complete and the interpolated image is ready to display. He quantified the cost for PC Gamer during the same Taipei trip: enabling XeSS MFG picks up roughly eight milliseconds of latency purely from that holding period. Hardware Unboxed’s 2022 testing found DLSS 3 frame generation raised total latency in Cyberpunk 2077 from 47ms to 63ms, a 16ms jump alongside a higher displayed frame count; Digital Foundry, testing concurrently, found DLSS 3 “inherently adds input latency due to the fact that it holds a frame.” Hardware Unboxed also recorded an 11ms penalty in F1 2022 under the same conditions.
NVIDIA has partially addressed this with Reflex, a system latency tool that reduces the CPU-to-GPU render queue depth, and more recently with Frame Warp inside DLSS 4, a reprojection technique borrowed from VR headset design that shifts output frames toward a player’s most recent inputs. Both reduce perceived latency without removing the underlying source, which stays tied to the pace of the base render loop.
AMD officially recommends a 60 FPS baseline before enabling FSR 3 Frame Generation. Below that floor, the latency additions aren’t masked and players feel every millisecond of the hold. On a gaming mouse in a reflex-heavy shooter, the extra 10 to 15 milliseconds is noticeable. On a controller in a single-player title, the same delay typically falls within the controller’s own inherent latency, and the player doesn’t register it.
Gaming handhelds sit at the center of that sensitivity gap. A device running at 35 watts rarely holds 60 native FPS in demanding titles at reasonable quality settings. Frame generation exists for those conditions, and the latency cost that comes with it doesn’t lift just because the screen is eight inches wide.
Intel’s Forward Gamble
What Changes with Extrapolation
Standard frame interpolation needs two rendered frames. Extrapolation needs one. It reads current hardware input alongside that single frame and predicts what comes next. The synthesized frame goes to the display while the GPU works on the genuine next render, eliminating the holding period that interpolation requires.
| Method | Frames Required | Latency Effect | Primary Risk |
|---|---|---|---|
| Interpolation (DLSS 3, FSR 3, XeSS MFG) | Two rendered frames before synthesis | Adds 8-16ms depending on hardware | Holds input responsiveness behind render pace |
| Extrapolation (Intel’s approach) | One rendered frame plus hardware input data | Designed to add zero additional latency | Misprediction creates visible frame artifacts |
How Petersen Described It
I’ve got one frame. I’ve rastered it. I’m showing it to a user. And while I’m not quite ready to raster a new one, I’m going to predict where he’s going to move his mouse.
Petersen, Intel’s graphics architecture fellow, gave that description to Gizmodo journalist Kyle Barr at a sit-down in Taipei on the eve of Computex 2026. What he outlined is a real-time input-prediction model embedded in the frame pipeline: velocity, direction, and acceleration from the hardware device go in; a synthesized next moment comes out while the GPU processes the genuine render behind it.
He also used the Gizmodo interview to place blame for player skepticism toward AI frame assistance. The reason gamers have become antagonistic toward the technology, he said, is that NVIDIA “fucked it up” during last year’s Multi-Frame Generation rollout. He served as NVIDIA’s director of technical marketing until 2019 before joining Intel.
Handheld Hardware, Where the Stakes Are Highest
The Arc G3 Extreme at Computex
Intel unveiled the Arc G3 Extreme at Computex 2026 alongside those interviews. Built on the Panther Lake architecture, it’s what Intel’s team called “a GPU SoC with a CPU inside it,” meaning the 12 Xe-core Arc B390 GPU takes the design priority and a 14-core CPU cluster works in support. Below 12 watts, the performance CPU cores park entirely, leaving efficiency cores to feed power to the GPU.
MSI’s Claw 8 EX AI+ is the first device carrying the chip. Acer’s Predator Atlas 8, also announced during the show week, uses Arc G3 and G3 Extreme processors in a comparable form factor with an 8-inch 1200p 120Hz display and an 80 WHr battery. Intel also introduced Endurance Gaming at the same briefing, a mode that dynamically caps frame rates and adjusts power; in Forza Horizon 6 at 1080p low settings, Intel claimed it extends battery life from under three hours to nearly six.
- 42%: Intel’s claimed average performance advantage over AMD Ryzen Z2 Extreme at 35W (Intel internal testing; independent benchmarks pending)
- 8W to 35W: Arc G3 Extreme’s configurable thermal design power range
- June 25: MSI Claw 8 EX AI+ launch date, priced at $1,499
Frame Energy Math on Portable Devices
One AI-generated frame costs roughly a quarter the energy of rendering an original, Intel’s team explained at the launch briefing. Four frames can be produced for the energy of about one and a half native renders. On a 35-watt thermal envelope shared between CPU and GPU, that ratio changes what’s achievable at a given battery level without requiring the full power that native rendering at equivalent frame rates would demand.
“You need frame gen to get a good experience” on handheld gaming PCs, he told PC Gamer in Taipei. Visual artifacts from frame generation are also less conspicuous on an 8-inch panel than on a large desktop monitor, which lowers the quality threshold the AI model needs to meet. For the category of compact gaming hardware pushing high refresh rates within tight power budgets, controller players on handhelds are the audience that already leaves frame generation disabled because of the latency cost. Extrapolation’s design goal is to remove that cost.
When Prediction Goes Wrong
The risk is a misprediction the player sees. If the model generates a frame showing the camera tracking right while the player has already moved left, the correcting real frame arrives moments later but the brief visual pop is there. He addressed this in the interview: “The question really is: how often do people change their choices, and how quickly can you respond to the change of choices?”
His position is that most player motion is continuous. The AI model, trained on real input sequences, handles straight lines, parabolic arcs, and sustained thumbstick holds cleanly. The failure case is what he called “a very notable physical change in your hand or your thumb”: a snap reversal, a sudden stop. On a controller, those inputs are rarer and less abrupt than on a gaming mouse. Handheld controllers already carry more than eight milliseconds of inherent signal latency in their communication chain, which means a brief misprediction can land within the margin players already absorb without perceiving it.
Intel hasn’t published prediction error rates, compute overhead, or end-to-end latency figures for the consumer version. The 2024 Intel research paper GFFE: G-buffer Free Frame Extrapolation for Low-latency Real-time Rendering, published by Intel engineers on arXiv, describes the engineering method but carries no consumer performance data. A rasterization trade-off also applies: the more AI frames the system generates per rendered one, the more GPU overhead the extrapolation model consumes from the same chip. On a handheld’s integrated GPU, that overhead competes directly with the render budget frame extrapolation is designed to supplement. How Intel distributes that cost across the Arc G3 Extreme’s power range hasn’t been specified publicly.
Desktop gaming with a gaming mouse is a different test entirely. Players in reflex-heavy competitive titles notice latency differences in the two-to-three millisecond range, well below the 8ms floor that interpolation adds to a handheld. Extrapolation introduces its own latency variable in the form of misprediction, and how often that occurs under competitive desktop conditions is one of the questions the technology hasn’t answered.
From Sydney to Taipei
Intel’s public record on frame extrapolation starts at SIGGRAPH Asia 2023, the graphics research conference held in Sydney that December. There the company introduced ExtraSS, a flow-based neural network that uses a game engine’s G-buffer data (depth, surface normals, geometry) to guide the extrapolation process. The method works where G-buffers are accessible, which excludes mobile platforms, cloud gaming pipelines, and forward-rendering engines that don’t expose that data layer.
The GFFE paper addressed that constraint directly. Intel published the work to arXiv in 2024, describing a recurrent neural architecture combined with optical flow and warping techniques that produce extrapolated frames without G-buffer access. The approach handles occluded regions and shading changes that simpler warping methods fail on, the kinds of artifacts that fast-moving elements like particle effects and transparent surfaces tend to reveal. “G-buffer Free” extends compatibility to engines and platforms that ExtraSS couldn’t reach, including mobile and cloud pipelines that most frame-generation research has historically skipped. The paper targets frame quality within what real-time rendering requires, though it carries no consumer-facing performance figures.
Between Sydney and Taipei, Intel stayed quiet on the consumer side. The Taipei interview was the first substantive public update in roughly two and a half years. PC Gamer’s reporter on the ground noted there had been a real possibility of a live extrapolation demo appearing at the show. It didn’t make the floor.
No release timeline followed the interview. The next milestone is a full public showcase with benchmark numbers, prediction accuracy data, and a working demonstration. Intel first flagged the technology in 2023. The demonstration hasn’t followed.
