Steam Deck Performance Guide

The Deck is a fixed target with the constraints of a budget gaming laptop and the cooling of a mid-range phone. Eight RDNA 2 CUs at up to 1.6 GHz is roughly 1.6 TFLOPs of GPU — a quarter of XSS, a sixth of PS5. There's no hardware ray tracing path; HWRT-Lumen and DXR materials silently fall back or fail. The CPU is four Zen 2 cores; UE5 game-thread budget on Deck is tighter than on any console.

And critically: the Deck is shipping now, with millions of units. Steam Deck Verified is a real cert process; titles that fail get demoted from "Verified" to "Playable" or "Unsupported", and players see the badge before they buy.

OLED isn't faster — same CUs, same clocks. The 6 nm process shrink yields better sustained thermals and battery; LPDDR5 is faster (~16% memory bandwidth bump). Don't ship a separate "OLED preset" expecting more frames — ship a "90 Hz target if you have headroom" mode that benefits OLED automatically.

A common community misconception is that the OLED Deck has a lower TDP cap. It doesn't — both LCD and OLED are configurable to 15 W via Steam OS settings.

This is the trap that catches every team's first attempt: UE's IsRunningOnPlatform macros do not return Steam Deck. Under Proton, the binary thinks it's running on Windows. Native Linux builds report Linux but not Deck specifically.

The right way to detect:

// Option 1: Steamworks SDK (preferred)
if (SteamUtils() && SteamUtils()->IsSteamRunningOnSteamDeck())
{
    // Apply Deck-specific scalability
    UDeviceProfileManager::Get().SetActiveDeviceProfile(TEXT("SteamDeck"));
}

// Option 2: Environment variable
FString DeckVar = FPlatformMisc::GetEnvironmentVariable(TEXT("SteamDeck"));
if (DeckVar == TEXT("1"))
{
    // Same
}

The Steamworks SDK approach is more reliable; the env var is set by Valve's launch wrapper but can be missing in development scenarios.

Configure your .uplugin to use LoadingPhase=EarliestPossible for the module containing this detection — the device profile must be applied before scalability tier evaluation, which happens early in startup.

Per Epic's Steam Deck Quick Start, the canonical structure is a Config/SteamDeck/ overlay plus a DeviceProfile entry:

[SteamDeck DeviceProfile]
DeviceType=Windows
BaseProfileName=Windows
+CVars=r.AntiAliasingMethod=4              ; TSR (or set 2 for FSR2 plugin)
+CVars=r.TemporalAA.Upsampling=1
+CVars=r.ScreenPercentage=70
+CVars=r.Lumen.HardwareRayTracing=0        ; Van Gogh has no HW RT
+CVars=r.Lumen.ScreenProbeGather.IntegrateDownsampleFactor=2
+CVars=r.DynamicGlobalIlluminationMethod=2
+CVars=sg.ResolutionQuality=80
+CVars=sg.PostProcessQuality=2
+CVars=sg.ShadowQuality=2
+CVars=sg.GlobalIlluminationQuality=1
+CVars=sg.ReflectionQuality=1
+CVars=sg.ViewDistanceQuality=2
+CVars=sg.EffectsQuality=2
+CVars=sg.FoliageQuality=1
+CVars=t.MaxFPS=30
+CVars=r.VSync=1
+CVars=r.SetRes=1280x800f
+CVars=r.DynamicRes.OperationMode=1
+CVars=r.DynamicRes.MinScreenPercentage=60
+CVars=r.DynamicRes.MaxScreenPercentage=85
+CVars=r.DynamicRes.FrameTimeBudget=33.3

The DeviceType=Windows is intentional — under Proton the Deck runs Windows binaries, and most Deck-shipped UE5 titles ship the Windows depot. Native Linux builds historically run slower than Windows-on-Proton (Triple-Aye reported 12–16 fps native vs 30 fps Proton on the same UE5 build).

Pick one before you tune anything else. Three reasonable choices:

• 30 fps locked at 15 W (cinematic UE5). The default for Lumen-on / Nanite-on UE5 ports. Hellblade II Enhanced, Robocop: Rogue City both target this.
• 40 fps for OLED 90 Hz. Half-rate of 90 Hz; viable on lighter UE5 content with FSR2 Quality.
• 60 fps with Lumen disabled. Possible on lighter Lyra-class projects; not viable on most Lumen+Nanite-shipped titles.

Pick at the device-profile level via t.MaxFPS. Pair with r.VSync=1 and Dynamic Resolution between 60% and 85% screen percentage. The 30 fps floor matches the "30 fps is sticky" community consensus on Deck-shipped UE5 titles.

Hardware ray tracing must be off — Van Gogh has no RT cores. The decision tree:

• Keep Lumen GI (SW): viable for Quality-mode 30 fps with aggressive tuning. Set r.Lumen.HardwareRayTracing=0, r.Lumen.ScreenProbeGather.IntegrateDownsampleFactor=2, r.Lumen.ScreenProbeGather.MaxRayIntensity=10, drop spatial filter passes to 1.
• Downgrade Lumen reflections: r.Lumen.Reflections.DownsampleFactor=2, r.Lumen.Reflections.MaxRoughnessToTrace=0.3.
• Kill Lumen entirely: r.DynamicGlobalIlluminationMethod=0 + r.ReflectionMethod=2 (SSR) for 60 fps mode. Bake static lighting; r.GenerateMeshDistanceFields=0.

For the deep mechanics see the Lumen Performance Deep Dive; the Lumen defaults gotcha covers the disable-it-properly recipe.

Keep Nanite on for high-poly hero meshes; the cluster culling pays off even on 8 CUs. Disable Nanite on:

• Foliage — software VRS path on Van Gogh is expensive; baked LODs win on Deck.
• Translucent / masked content — programmable raster cost is amplified on weaker silicon.
• Distance-rendered set-dressing — HLOD or low-poly proxies beat Nanite at low screen-pixel sizes.

Hellblade II Enhanced demonstrates the workable case: 30 fps, FSR2 Balanced, full Nanite + SW Lumen, mostly-baked indirect lighting on Deck. They got Steam Deck Verified status post-launch, with frame-time spikes acknowledged.

FSR is the right upscaler for Deck — AMD's own algorithm, tuned for RDNA 2, no NVIDIA hardware required. Setup:

• Install AMD's FSR3 UE Plugin from GPUOpen.
• Set r.AntiAliasingMethod=4 (TSR) AND r.TemporalAA.Upsampling=1 — the FSR plugin requires Temporal Upsampling enabled. (TSR's own internals get bypassed when FSR3 takes over.)
• r.FidelityFX.FSR3.Enabled=1, r.FidelityFX.FSR3.QualityMode=2 (Balanced; 1.7× scale = 753x470 input at 1280x800 output).
• Avoid Frame Generation on Deck — the 30 fps floor means frame-gen would target 60 fps from a 30 fps base; latency cost is too high for the visual gain.

FSR2 still works and may be more performant than FSR3 for a 30 fps locked target. FSR3 plugin requires UE 5.1.1+; 5.1.0 has an ABI incompatibility.

• Shader compilation hitches on first run — the DXVK/VKD3D state cache is empty. PSO precaching (UE 5.2+) is essential; see the PSO Precaching tutorial. Hellblade II at launch suffered badly here.
• EAC SDK on Linux — if you use Easy Anti-Cheat, the Linux .so swap is required for SteamOS. Test specifically; cert-equivalent failure mode.
• DefaultGraphicsRHI — some UE5 titles ship DX11 RHI on Deck for stability under Proton. Modern UE 5.5+ Vulkan-on-DXVK is generally fine; test both.
• Wayland clipboard / overlay quirks — Steam overlay can interact poorly with custom UE5 input handlers on Wayland. Test the overlay-on-overlay-off pattern.
• Immutable filesystem — SteamOS is read-only by default. Don't write outside the prefix or save folder.

Tools that work on Deck:

• MangoHud — the Deck-equivalent of Afterburner; CPU/GPU/RAM overlay. Free, ships with SteamOS.
• Unreal Insights over network — launch dev kit and target Deck IP via -tracehost.
• stat unit / stat gpu / DumpGPU in a Development build.
• RGP captures on a Deck-class RDNA 2 GPU on a desktop machine, transferred over — the patterns translate (occupancy, async overlap).

Before submitting for Steam Deck Verified review:

1. Default settings are Deck-tuned out of the box — the Verified bar is "playable at recommended settings" without the player tinkering.
2. Controller glyphs and Steam Input — Verified requires the right glyphs surface, including for any non-controller HUD prompts.
3. 800p default with Dynamic Resolution band 60–85%.
4. 30 fps locked with t.MaxFPS=30 in the Deck device profile.
5. FSR2 Balanced as the default upscaler.
6. HWRT explicitly off — r.Lumen.HardwareRayTracing=0.
7. PSO precaching enabled with bundled cache from QA playthrough — first-run hitches are an instant Verified rejection signal.
8. Sleep/resume tested — close the Deck lid, open it, verify the game resumes correctly.
9. Battery-life test — 90 minutes of gameplay at 15 W; if you're throttled below 30 fps mid-test, your TDP target is too aggressive.
10. OLED + LCD both tested — the LPDDR5 bandwidth difference is small but real on memory-bound scenes.

The Steam Deck is a uniquely good fit for PerfGuard's baseline workflow because the hardware is identical across millions of units — unlike PC, where regressions get masked by GPU variance. A single PerfGuard run on a dev Deck applies to every player.

• Upscaler Tuning — FSR setup and quality regression checklist.
• Lumen Performance — the SW Lumen tuning the Deck profile uses.
• PSO Precaching — mandatory for surviving the first-run experience on Proton.