v2.1.0

GFG Extreme 2.1.0 — Frame Fusion Core preview

Experimental prerelease. The stable release remains 2.0.3.

Frame Fusion Core can share compatible real or generated frame buffers with the MAKO scaler, removing one redundant source-image copy while preserving the existing fallback. It supports native x86_64 SDR with the open MAKO scaler and an actual resolution split. The feature defaults off and is enabled in Settings → Diagnostics → Frame Fusion; install the updated engine and restart the game.

The package includes a source-built native renderer fork, immutable upstream and patch provenance, native image/copy/synchronization checks, and a standalone GPU proof tool. UI and exported diagnostics distinguish requested settings from observed processing. No whole-device energy or Steam Deck FPS improvement is claimed yet.

Flatpak, 32-bit, LS1 and HDR use their previous paths. Shared LSFG motion analysis and full Frame Fusion are not part of this preview. Physical Deck and game validation remains required.

SHA-256: e8db3fdcc5709a9654b49d7bf0579aea82f0675f73e90566b5e53c2bfd597e6e

GFG Extreme 2.0.3

GFG Extreme 2.0.3

Fixes control conflicts found in the original Steam Deck log recorded on 2.0.1 at 2026-10-12 00:34:41.

• Autopilot now separates actual renderer/power transitions from ordinary consumed Budget windows. Healthy eight-second baselines can reach opted-in Half Rate Shading and GPU clock trials; real settings changes, Steam menus, stale telemetry and concurrent resource experiments still block admission.
• Output-goal decisions use independent evidence windows. Collecting the next window no longer erases preceding misses. Sustained shortfall can trigger the existing achievable-goal policy, and display changes remain honored during collection.
• Emergency TDP increases remain temporary. Once safe measurements permit a power search, Auto's ordinary HOLD and new clock/shading trials cannot freeze emergency watts. Lower-cap failures restore the confirmed working setting.
• Frame OS Act cannot take or keep the adaptive overlay at an emergency cap in the economy modes. Normal Act yields for power research every five minutes instead of ten. Extreme continues to honor the user's performance ceiling.
• An explicit game override of render resolution is recognized before requesting scaled points. The refusal is remembered for that game; resetting learned data allows a new trial without immediately replaying the old refusal.

Fourteen focused regressions accompany the existing software checks, including actual loop execution, actuator admission, TDP rollback, display/menu protection and scaling refusal. The release workflow runs backend, native Vulkan/HUD and frontend checks and verifies the Decky archive before publishing.

The source recording tested Balanced and Autopilot Auto on 2.0.1, not Battery or Quality on 2.0.2. These fixes retain the 2.0.2 mode objectives. They do not promise a specific minimum wattage or measured battery/FPS gains. The upstream renderer is unchanged; a game that ignores spatial scaling is not forced into an unsupported resolution.

Install using Decky Loader -> Install from zip -> GFG-Extreme-v2_0_3.zip, then restart the game.

SHA-256: 54b28e61d4d7673f521a69e27f2dafa48a3aed2bc310b35a87fc91a690604b09 (also in SHA256SUMS.txt).

GFG Extreme 2.0.2

GFG Extreme 2.0.2

Official mode-objective update for Battery, Balanced and Quality.

• Battery now targets the lowest verified feasible TDP, rather than keeping a fixed 6 W comfort floor when the device reports a lower minimum. After finding a floor for the current real cadence, it may confirm a cheaper compatible normal cadence/scale and resume searching down. Failed trials restore the known-good point and watts; unavailable ratios and ignored power caps cannot claim savings. Autopilot's Battery preference uses the same minimum-power goal.
• Balanced keeps its 30-real-FPS floor and normal power ceiling, and can retest a finer frame ratio when power cannot be reduced further. Heat and rejection cooldowns block these trials.
• Quality now preserves saved render resolution in both recommendations and the actual trial ladder. After five minutes of clean play, it can trial a finer ratio at the saved power budget, with the previous point retained for rollback. Failed points cool down for ten minutes instead of remaining rejected forever within the same session.
• The interface describes Battery's minimum-power goal and Quality's resolution/recovery behavior.

Sixteen focused objective tests accompany the existing mode, telemetry, restoration, replay, native Vulkan/HUD and frontend tests. See docs/MODE_OBJECTIVES.md. The release workflow repeats software checks and validates the Decky archive before publishing.

Model-based test examples (7 W versus 6 W, or 3 W on a sufficiently light scene) are not measured improvements on a physical Steam Deck. Actual minimum power depends on the game, target cadence, renderer capacity and hardware. A failed hardware-minimum probe returns to the proven feasible value. The bundled upstream renderer is unchanged.

Install using Decky Loader -> Install from zip -> GFG-Extreme-v2_0_2.zip, then restart the game.

SHA-256: 37ad08427e229d658ca3a2b6b01fb6962bb046be974421608365c0b08911c742 (also in SHA256SUMS.txt).

GFG Extreme 2.0.1

GFG Extreme 2.0.1

Official maintenance release for Autopilot and Frame OS control paths.

Fixes

• GPU Clock and opted-in Half Rate Shading can now be proposed on stable, heavily loaded game cadence. Previously they were proposed only below the FPS goal, while admission required at least 95% of the goal: every shading request in the supplied 2.0.0 extract was skipped. Confirmation, clean baseline and Steam-menu guards remain required.
• SteamOSManager's manual GPU clock step starts below the observed GPU MHz, rather than below the hardware maximum, so the proposed downward step cannot accidentally pin a lightly loaded GPU at a higher frequency.
• GPU/VRS verification times out on repeated or stale samples, checks heat immediately, and requires post-change windows. The core loop reaches this timeout even when the renderer becomes silent.
• Pending GPU/VRS restoration blocks new resource control across Battery, Balanced, Quality and Autopilot. An idle GPU restore cannot clear a shading restore lock. Plugin restart restores an unjudged clock trial; QAM overrides remain respected during retry and while holding a verified clock.
• Skipped/rejected actuator trials leave the ordinary power Governor available. Failed writes back off; the HUD does not announce an experiment after it has ended.
• Steam menu samples, resumed pre-transition windows and stale telemetry cannot lower the Autopilot FPS goal.
• Repeated short Frame OS boosts that never show increased real cadence now accumulate failure exposure after the interval-ring settle period. Backoff releases the adaptive executor point to Governor rather than holding it indefinitely. Missing cadence alone is not learned as failure.
• Point release restores GFG-owned GPU/VRS settings before returning to the base rendering point.

Evidence and validation

Audited diagnostics/autopilot-2.0.0-20261011-230354, commit c3bd63e: 2,184 session samples over 2,206 seconds, median 29.999 real / 89.993 output FPS, and 15 verbatim skipped-shading events. The extract reports unchanged 30 real FPS in boost and calm. All 15 shading events are preserved in a regression fixture. See docs/RELEASE_AUDIT_2_0_1.md for evidence boundaries and test coverage.

Backend, native Vulkan mock-driver/HUD tests, frontend build/smoke and Decky archive integrity/version checks must succeed before this workflow publishes the ZIP.

Limits

The diagnostics branch contains a summary and skipped events, not the full raw recording. These fixes are software-validated; GPU/VRS behavior and performance gains in 2.0.1 still require a physical Steam Deck test. No increase in unique display frames, input-to-photon latency or battery life is claimed. The upstream bundled renderer is unchanged; generated-image acquisition timeouts are not claimed resolved. The logged game overrides render extent, so render-scale reductions cannot make that game render at a lower resolution. Half Rate Shading stays opt-in and can reduce clarity. Display refresh and Allow Tearing are not changed by these fixes.

Installation: Decky Loader -> Install from zip -> GFG-Extreme-v2_0_1.zip, then restart the game.

SHA-256: 18b04141c3bf76b927e6860212d52b3fc5cb4b3c7ad0d3ec656b9664a7b81ff9 (also in SHA256SUMS.txt).

GFG Extreme 2.0.0

GFG Extreme 2.0.0 | Autopilot

Official release • 11 October 2026

Autopilot replaces the former Extreme mode. An existing Extreme profile is migrated without overwriting its Saved rendering settings. The normal Governor remains the sole TDP controller.

Main functionality

• Real/output FPS targets based on renderer evidence; automatic 90/60 goal selection when measured cadence cannot maintain the panel goal.
• Governor adjusts the verified TDP limit. Failed lower-watt experiments roll back and are remembered; repeat failures now back off for up to 30 minutes (including across restarts).
• SteamOSManager GPU Clock trials are restricted to likely CPU bottlenecks and respect Steam QAM manual overrides, with verified readback and rollback. Unsupported devices fail closed.
• Opt-in Half Rate Shading (RADV VRS 2x2) experiments have exclusive ownership, verification and restore. They are not attempted while the Steam Quick Access Menu is open or before a stable, confirmed operating point and fresh evidence.
• Autopilot's HUD uses measured real/output FPS, GPU and CPU activity, temperature, frametime and APU watts. Deferred, blocked and failed experiments are now reported as skipped or paused, never as successfully applied.
• Stop, game exit and mode changes restore GFG-owned GPU/VRS settings or retain a retryable restore-pending lock.

Fixes verified from the 11 October field log

• Lower watt failures such as the repeatedly unsuccessful 9 W trials no longer trigger frequent repeated dips without a significant new opportunity.
• UI no longer states that GPU Clock or Half Rate Shading is being tested if the experiment has not started.
• Plugin, Governor and diagnostic exporter now all identify this build as 2.0.0 rather than mixing 2.0.1 and 2.0.4.
• The Steam QAM transition can no longer be mistaken for a clean performance comparison.

Validation and limitations

Automated backend tests, native Frame OS/HUD Vulkan mock-driver tests, frontend smoke tests and Decky archive integrity checks run in CI before publication. No gain in unique display frames, input latency, image quality or battery life is claimed from the available session.

GPU-clock writes and dynamic Half Rate Shading have not yet been validated on a physical Steam Deck LCD/OLED in this release. The opt-in shading feature may reduce text/HUD clarity. The renderer's occasional generated-image acquisition timeouts and Frame OS Act responsiveness have not been proven resolved by this update. Allow Tearing is not changed automatically.

Installation: Decky Loader → Install from zip → GFG-Extreme-v2_0_0.zip, then restart the game.

SHA-256: e838f90aa08b7f415ddf7eee70d000acfc5e2f47a3b3fb82faa38aaeae1fd7ad (also in SHA256SUMS.txt).

GFG Extreme 1.6.9

GFG Extreme 1.6.9

Official release of the ordinary GFG renderer. Restart running games after upgrading.

Decisions use delivered frames

• A diagnostics line that only carries a planned multiplier is no longer a frame the Governor can act on. Power steps, the guard, Act's starvation check and the session average ignore it. Measured output, including a measured zero, and scheduler-interval samples still count.
• The text HUD no longer prints that planned rate as FPS. The ring HUD already refused it.
• A rejected operating point in a session report now shows the measured output and real cadence next to what was requested, when that evidence exists. A planned multiplier is not described as delivery.

Unchanged

Extreme still clamps both PPT channels to the inherited ceiling and keeps that clamp across a new session, a missing overlay and a telemetry gap. The HUD still publishes twice a second. Flow trials, Battery, Balanced and Quality keep their existing goals. Saved profiles and the generator model are unchanged.

Validation and limits

Backend regression tests cover a plan-only window, a mixed window, the text HUD and the rejection report. Native Frame OS/HUD integration, HUD pixels, frontend smoke cases and the final Decky archive are checked before publication.

Extreme, Frame OS Act and automatic flow tuning remain experimental features inside an official release. GFG Open remains retired. This release stops planned FPS from steering power; it does not claim a measured on-device FPS, latency or ghosting improvement.

GFG Extreme 1.6.8

GFG Extreme 1.6.8

Official release of the ordinary GFG renderer. Restart running games after upgrading.

Keep optimization accountable to delivered frames

• Accepted automatic flow-scale choices continue to monitor fresh cadence. A severe shortfall restores the player's Saved flow value immediately; two consecutive moderate shortfalls also restore it. Missing or changed applied flow state removes the success claim and triggers an acknowledged rollback. Failed choices are not retried in the same context.
• Incomplete frame-generation state supersedes an earlier flow acknowledgement. An old successful value can no longer stand in for a newer missing value or unavailable resources.
• Spatial-layer diagnostics no longer replace frame-generation capacity or game FPS. This keeps unrelated zero-slot reports from disabling the ordinary generator or blocking Extreme's deeper ratios.
• Measured zero output, when real cadence is available, stays zero instead of becoming a theoretical 90 FPS from the planned multiplier. Plan-only samples cannot prove a flow benefit and do not populate the Rings HUD's live FPS counter.

Existing controls and release packaging

Extreme retains the quarter-step ratios through 4x, subject to actual renderer capacity and cadence checks. Battery, Balanced and Quality keep their separate optimization goals. Saved profiles, the generator model and the inherited power ceiling are unchanged.

ENERGY retains saved TDP allowance in its centre and battery charge in its arc. TDP remains green on external power and changes with battery charge when unplugged. The theme installer remains available in Settings.

The startup log now identifies the installed version correctly. Repeated builds can reuse a cached previous-release archive instead of downloading it again; archive checksums and ZIP CRC are still verified on every build. Initial cache misses still count as GitHub asset downloads.

Extreme ceiling and HUD rate

• While Extreme stays enabled, a new game session, a missing overlay or a gap in FPS telemetry keeps both PPT channels at the clamped ceiling. Those paths used to restore the higher inherited fastPPT for one Governor tick. Turning Extreme off, changing profile or any real exit still restores the original caps.
• A failed power claim outside Extreme is no longer labelled extreme-ceiling-not-applied.
• The in-game HUD is published twice a second, the same rate the Vulkan layer already polls. Governor power decisions stay once a second. The MangoHud config file is still rewritten at most once every 5 seconds.

Validation and limits

Backend regression tests cover held-flow starvation, duplicate samples, healthy recovery, unexpected and incomplete acknowledgements, role isolation, measured zero output, plan-only HUD data, real overlay rollback, Extreme ceiling handoff and the half-second HUD publish. Native Frame OS/HUD integration, HUD pixels, frontend smoke cases and the final Decky archive are checked before publication.

Extreme, Frame OS Act and automatic flow tuning remain experimental features inside an official release. GFG Open remains retired. These fixes prevent invalid optimization and misleading evidence; they do not claim a measured on-device FPS, latency or ghosting improvement.

GFG Extreme 1.6.5

GFG Extreme 1.6.5: bounded transitions and reversible runtime foundations

Changes

• Extreme claims power with both fastPPT and slowPPT already bounded by the stock 15 W ceiling or a lower inherited user limit. Renderer trials stop if applying the ceiling fails.
• Internal mode/display replanning keeps the Extreme ceiling instead of temporarily restoring a previous 20 W cap. Leaving Extreme restores inherited settings.
• Successful ceiling retries resume the search; an old failure status no longer blocks recovery.
• Rejected operating points record fresh requested and delivered real/output FPS, multiplier, renderer acknowledgement, sample count and observation span. A reused telemetry session cannot supply evidence for the request.
• Native Frame OS gains an optional stall-shield policy and explicit support/active telemetry, preserving the existing behavior when the policy is off.
• The privileged helper gains bounded leases and persistent undo for GFG-marked game thread priority, OOM preference and OEM fan-controller selection. PID/TID identity checks, readback, EOF/expiry recovery and external-change preservation are covered by regression tests.
• Updated launch wrappers identify GFG-managed games for scoped resource control.

Current scope

The resource primitives and optional shield are foundations for the remaining booster integration; they are not automatically activated or represented as completed boosters. OOM preference is not swap/huge-page tuning. OEM controller selection is not predictive cooling.

This release does not claim a fix for Ghost of Tsushima ignoring render scaling, all fractional-point failures, Act starvation or ghosting. delivered-deeper-ratio can mean the renderer maintains target output using fewer real frames; the new evidence makes that distinction inspectable.

Validation

GitHub Actions checks backend regressions, native Frame OS/HUD tests and ABI, frontend build/smoke tests, HUD cost and release packaging. Device-level FPS gain, controller behavior, input latency and image quality still require a physical Deck comparison.

Install

Download GFG-Extreme-v1_6_5.zip and install through Decky Loader (Install from zip). Relaunch the game to use the updated launcher and native layers.

SHA-256: a0301d163b397169f02d7f9df8a8ccf5a7c27118729f7079aed3519be8a59e28 (also in SHA256SUMS.txt).

GFG Extreme 1.6.4

GFG Extreme 1.6.4: trustworthy audit summaries

The three user-recorded sessions on 1.5.0–1.6.1 exposed omitted transition samples and inherited A/B results. This release fixes diagnostics so they no longer hide those cases.

Fixed

• Extreme summaries inspect both fast and slow PPT channels, including APPLY transitions without a confirmed operating point.
• Reports distinguish all observed caps from caps observed while GFG owns control, and count above-ceiling samples against each sample's applicable ceiling. Readback alone is not attributed to a particular writer.
• Frame OS records separate current-session and historical A/B statistics. Reports identify historical-only and mixed old results instead of describing them as fresh measurements.
• Preliminary measurements are labelled when the record provides that confidence flag. The response metric is explicitly internal pacing/freshness telemetry, not independently measured input latency.
• Relay input with an unknown device count is no longer reported as zero controllers. Per-decision sample counts are labelled as samples rather than seconds.

Scope

This release improves the evidence used to diagnose Extreme. It does not claim to fix Act output starvation, deliver render scaling in Ghost of Tsushima, remove ghosting, or prove that transition overshoots are resolved on a physical Deck. Historical archives cannot yield a current-session A/B mean when only a mixed mean was stored. Runtime control and frame presentation are unchanged; the A/B meter adds small statistics to its existing status snapshots.

Validation

Regression coverage includes missing-point transition samples, fast-only cap violations, unowned reads, changing ceilings, historical-only/mixed/current A/B provenance and unknown relay device counts. GitHub Actions runs backend, native Frame OS/HUD, frontend smoke/build, HUD benchmark and release archive checks.

Install

Download GFG-Extreme-v1_6_4.zip and install through Decky Loader (Install from zip).

SHA-256: d8df4f55989f827f14b76b91be1d2259e9c47f4d759db6cf95cfcedcbf0fd864 (also in SHA256SUMS.txt).

GFG Extreme 1.6.3

GFG Extreme 1.6.3: power split that always keeps its undo

Fixed

• Partial CPU caps restore correctly. If one CPU policy accepted a lower clock and another failed, the undo previously depended on a cap that had not yet been committed. GFG now detects the actual changed values and restores them.
• Recovery survives a failed recovery. Busy or unreadable CPU policies keep their undo entries for retry. A write must read back correctly before recovery is complete; GFG does not learn its own leftover cap as your original clock limit.
• An outside CPU change does not strand the other policies. Your new value is preserved, while the remaining GFG-owned caps return to their original values.
• No CPU write without its undo record. If the recovery marker cannot be saved, the optimization is skipped. Journals keep both the planned cap and the previous value so a crash halfway through the next cap can recover both.
• The helper preserves the current owner's clock limit. A new ownership session rebases the original value after an external change. CPU restore is now a separate helper operation: helper shutdown cannot put a recovered old cap back.
• SteamOS Manager checks both PPT channels. A half-applied pair is not accepted. Whole-watt rounding never raises a fractional ceiling; GFG uses the direct pair writer when available, otherwise the lower permitted whole-watt step.
• Exact restoration is reported honestly. A manager-only system that cannot represent the original fractional or unequal caps reports the unsupported restore rather than claiming success at different values.

Status and compatibility

Power Split shows a pending CPU restore and retries it before another optimization. Existing settings and profiles are preserved. Normal successful control keeps the same polling cadence; no new work is added to the frame-presentation path.

Validation

Regression coverage includes partial writes, crashes between cap changes, failed and unverified recovery, unavailable journal storage, external changes, late helper writes, helper ownership sessions and shutdown, fractional PPT ceilings and both-channel readback. Backend, native Frame OS/HUD, frontend, HUD benchmark and archive checks run in GitHub Actions. Physical Deck validation is not claimed.

Install

Download GFG-Extreme-v1_6_3.zip and install it through Decky Loader (Install from zip) over your existing installation.

Verify your download

SHA-256: 98af07d12554c2157160995fadb01ca26d07ea83cf8637e207e6f058ad11a32e (also in SHA256SUMS.txt).