Skip to content

Fields

On-device & performance latest news

Running AI on players' devices: on-device models, VRAM and frame-rate budgets, neural rendering and upscaling.

37 picks21 in the last 30 days56 collected in all

Updated

Picks archive · Page 2

Oct 26Mon21–37
  1. arXiv Game World Models & Generation56

    WorldPager: zero-stall tiered memory for long-horizon world models on a single consumer GPU

    WorldPager stores the int4 KV memory of interactive world models across GPU, DRAM and NVMe tiers and prefetches it based on camera pose. On a single RTX 4090, it continuously generates 37.5 minutes of video without stalls, with host memory staying constant and a speed of about 9.5 frames per second. On a closed test set, int4 storage showed no measurable quality loss. On long-interval revisits, image quality dropped by about 1 MUSIQ point.

    Why it matters: The paper places a world model's KV memory across GPU, DRAM and NVMe tiers with pose-based prefetching, which is useful for readers following long-horizon interactive video generation.

Oct 23Fri
  1. Unity Blog65

    Unity MethodReload skips domain reload, experimental feature speeds up iteration

    Unity has introduced the experimental MethodReload feature. Developers only need to mark methods and start the watcher, and edits to method bodies take effect immediately in a running Editor or connected Player, with no need to wait for a domain reload.

    Why it matters: The article explains the causes of domain reloads and the boundaries of MethodReload's use, helping readers judge whether it fits their editor or on-device iteration workflow.

Oct 22Thu
  1. GamesBeat58

    Wonder Interactive launches Strata, using AI to turn Unreal Engine projects into browser-playable versions

    Wonder Interactive has launched Strata, an AI development tool that turns Unreal Engine projects into browser-playable links using WebAssembly and WebGPU.

    Why it matters: It lays out the concrete workflow from project to web playtest link, along with browser test results for City Sample, which helps readers judge whether a web-based playtest approach suits their own projects.

  2. GamesIndustry.biz35

    AMD's FSR 4.1 upscaling is coming to Xbox Series X|S

    AMD and Microsoft have announced that AMD's FSR 4.1 machine learning upscaling technology will support Xbox Series X|S consoles, but for now they only say it is "coming soon" and have not given a specific release date. Developers will need to patch their games to enable FSR 4.1. Older titles will not be upgraded automatically. AMD has not disclosed other details about the Series X|S version. According to GamesIndustry.biz, the Series X|S lacks dedicated machine learning hardware, so the gains may fall short of the PSSR and QSSR upscaling on Sony's PS5 Pro. The same announcement says FSR 4.1 will also support Radeon RX 6000 series graphics cards, Gorgon Halo and Strix Halo Ryzen 9 laptops, and handhelds such as the ASUS ROG Xbox Ally, which use the Z2 Extreme and Z2A chips.

    Why it matters: FSR 4.1 is worth comparing with PS5 Pro because the article notes that Series X|S lacks dedicated machine learning hardware, which may make its console-side upscaling gains smaller than PS5 Pro's.

Oct 21Wed
  1. Inworld AI39

    Latitude uses Inworld to add full character narration to Voyage

    Inworld provides voice synthesis and model inference for Latitude's game Voyage, so every story beat can be narrated with character voices. In the case study, voice costs are about a quarter of Latitude's previous provider's, and TTS runs 3–5 times faster than another provider in Latitude's production tests. Over six weeks, the platform handled more than 15 million TTS calls and 25 million inference calls.

    Why it matters: The case study offers vendor-disclosed figures for voice costs, TTS speed, and call volume, which can be used to weigh the cost and infrastructure trade-offs of fully narrating a game.

Oct 20Tue
  1. arXiv Game AI69

    Matrix-Game 4.0 harnesses video models to build real-time playable game worlds

    Matrix-Game 4.0 separates game execution from visual generation. A lightweight game engine maintains the world state and executes interaction rules, while an autoregressive video model turns low-polygon rendered frames into detailed images. The paper reports real-time 480p gameplay at 60 FPS on a single NVIDIA RTX5090, with a key-press-to-display latency of 112 ms. Single-step generation sustains 83 FPS throughput, with peak GPU memory usage of 18.2 GiB.

    Why it matters: This work hands world state and interaction rules to an executable engine and leaves image generation to a video model, offering a useful reference for the engineering split in playable world models.

Oct 19Mon
  1. Youxituoluo46

    How much would it cost to have game NPCs remember players?

    The article assumes 100,000 monthly active players with 100 dialogue rounds per player per month, and uses public large model API prices to estimate that model inference and memory processing would cost about US$6,908 per month. This excludes costs such as voice, servers, and operations. The article notes that NPC memory should be split into recent raw text, summaries of events and preferences, and on-demand retrieval, and that costs grow roughly linearly with player numbers.

    Why it matters: The article breaks down costs based on an assumption of 100,000 monthly active players with 100 dialogue rounds per player each month, and compares the costs of cloud APIs with those of on-device deployment. It can serve as a reference for estimating your own project.

Oct 16Fri
  1. GameLook42

    Epic rebuilds EGS launcher with AI coding, claims store pages load 46% faster than Steam

    Epic released a preview of its new EGS launcher on October 14, saying it was rebuilt with AI "vibe coding." Startup time is cut in half, and cold-starting into the store page is 46% faster than Steam, though these are Epic's own test results. The new version supports Windows and macOS, adds a Linux version, and can bring games from other PC game platforms into a single launcher view.

    Why it matters: The article sets the launcher's performance figures alongside Steam's AI disclosure rules, offering a look at Epic's calculations in platform competition and its engine ecosystem.

Oct 15Thu
Oct 14Wed
  1. Inworld AI Blog49

    Inworld launches Inference service for consumer AI

    Inworld has launched Inworld Inference, an inference service built for consumer AI applications and now open to all developers. The company says that on p95 time to first token, Gemma 4 26B is 7.8 times faster than the median OpenRouter provider, with input priced at $0.07 per 1M input tokens.

    Why it matters: The article breaks down the latency and cost drivers of a conversational AI turn, from time to first token and caching to routing, and can be used to compare inference service options for conversational AI.

Oct 9Fri
  1. arXiv Game AI43

    LiveSynth: A Streaming Neural Synthesizer for Instrument Cloning and Text-to-Instrument

    LiveSynth is a streaming neural synthesizer that supports zero-shot instrument cloning and text-to-instrument generation, and can be played interactively in real time. It uses a feedback-free causal Transformer conditioned on MIDI events and CLAP timbre embeddings to output continuous audio latents at 100 Hz, which a causal variational autoencoder then decodes into audio; frame-level timbre conditioning enables continuous timbre morphing, and MIDI conditioning dropout allows autonomous continuation from a performance prefix.

    Why it matters: The paper sets out the architecture and real-time performance data of a streaming neural synthesizer, useful for understanding low-latency generation for playable instruments.

Oct 7Wed
  1. arXiv Game World Models & Generation39

    CtrlCache: Training-Free Control-Aware Caching Accelerates Interactive Video World Models

    The authors propose CtrlCache, a training-free control-aware caching framework for accelerating chunk-by-chunk autoregressive generation in interactive video world models. It exploits the fact that control signals arrive before denoising, labeling each video chunk as initial, transition, turning, or steady-state; full computation is retained only for initial and transition chunks, while turning and steady-state chunks reuse the transformer residuals from the most recent full computation step within the same chunk, and a frequency-mixed historical prior reuses the low-frequency structure of the previous clean latent.

    Why it matters: The paper uses control signals ahead of time for cache reuse decisions, achieving 1.21x to 1.41x inference speedup on three world models without training.

Oct 6Tue
  1. arXiv Game AI35

    Emoception proposes the SALFT framework, selectively fine-tuning a video ViT to recognize changes in player arousal

    The paper proposes SALFT (Selective Affective Layer Fine-Tuning), a framework that fine-tunes a Video Vision Transformer using a selection criterion based on layer-wise parameter L2 norm change to recognize changes in player arousal from gameplay footage.

    Why it matters: The paper uses layer-wise parameter L2 norm change as its selection criterion and shows that updating only about 8% of parameters approaches full fine-tuning, useful for comparing compute costs against your own player emotion recognition pipeline.

Oct 3Sat
  1. arXiv Game AI55

    CurveCodec 2 introduces skeleton-agnostic animation compression, encoding residuals with a learned entropy model

    CurveCodec 2 is a skeleton-agnostic codec for compressing skeletal animation. It encodes each sub-track as a curve in the log map and entropy-codes the residuals with a small learned entropy model, and its integer inference is bit-exact across platforms.

    Why it matters: The paper compares against ACL, a production-grade animation compression library used in game engines, clip by clip on 4,472 held-out test clips, and reports compression ratios and single-core decoding conditions under different error contracts.

Oct 2Fri
Sep 23Wed
  1. NVIDIA GameDev39

    NVIDIA announces this month's updates for RTX game developers, including DLSS 5, ACE upgrades and RTX Kit 2026.3

    NVIDIA announced this month's updates for RTX game developers, including DLSS 5 with 3D-Guided Neural Rendering controls, speech pipeline extensions and inference framework updates for NVIDIA ACE, and RTX Kit 2026.3, which includes RTX Mega Geometry 2.0. The official tweet links to the developer blog post for this update; capabilities and integration details are in the original on developer.nvidia.com.

    Why it matters: The official post lines up DLSS 5, NVIDIA ACE and RTX Kit updates side by side, so developers can compare integration priorities and upgrade order for their own projects.

Aug 25Tue
  1. NVIDIA Blog38

    EA, Ubisoft, Embark and other publishers bring PC games and anticheat support to NVIDIA RTX Spark

    NVIDIA announced at Gamescom that publishers including EA, Embark and Ubisoft will bring games and anticheat support to RTX Spark, which launches this fall, with EA Javelin Anticheat also natively supporting the platform.

    Why it matters: The partner list ahead of RTX Spark's fall launch grows from May's COMPUTEX to this Gamescom, so readers can compare the two batches of publishers to gauge how adoption of the new platform is progressing.