Device profiles
Match GPUs by vendor, name, device id, VRAM, driver and platform to cap the quality tier and set cvars, block known-bad drivers, and size streaming to VRAM.
A device profile is a rule that recognises a GPU and adjusts the engine for it: it caps the quality tier the auto-detection may pick and sets console variables. It is Mowo's form of Unreal Engine's DeviceProfiles.ini. The same files carry a driver blocklist that names driver versions known to break the engine and tells the player what to do.
Where the rules live
Rules are INI sections in the DeviceProfiles config hierarchy, lowest layer first:
Engine/Config/BaseDeviceProfiles.ini: the rules the engine ships.Engine/Config/<platform>/<platform>DeviceProfiles.ini.- Your project's
Config/DefaultDeviceProfiles.ini. Config/<platform>/<platform>DeviceProfiles.iniin your project.
A project adds a rule with a new section, changes one by writing a section of the same name, and switches one off with Enabled=False.
; The 3050's 8 GB holds Epic textures, but this game's Epic GI is too heavy
; for it: cap it at High and keep the GPU-hot alert a little lower.
[DeviceProfile MyGame_RTX3050]
Description=GeForce RTX 3050
Priority=20
Vendor=NVIDIA
+NameContains=RTX 3050
+DeviceIds=0x2507
+DeviceIds=0x2582
TierCap=High
+CVars=telemetry.Alert.GpuTempC=84How a rule matches
| Key | Matches when |
|---|---|
Vendor | the PCI vendor is NVIDIA, AMD, Intel, Apple, or the id you write (0x10DE) |
+NameContains | the GPU name contains the text, case-insensitive |
+DeviceIds | the PCI device id is the id or inside the range (0x2200-0x25FF) |
SteamDeck=True | the platform reports Steam Deck hardware |
+NameExcludes | the name does not contain the text |
MinVramMB, MaxVramMB | dedicated video memory is inside the bounds (an unknown size never meets MaxVramMB) |
MinDriver, MaxDriver | the vendor driver version is inside the bounds, compared field by field (566.36) |
Integrated | True for integrated or unified-memory GPUs only, False for discrete only |
+Platforms | win-x64, linux-x64, osx-arm64, or a family prefix (win, linux, osx) |
+GraphicsApis | D3D12, D3D11, Vulkan or Metal |
NameContains, DeviceIds and SteamDeck are alternatives: a rule can name a GPU family by marketing name and by PCI id range, and a device that reports either matches. Every other key that is set must hold. Among the rules that match, the highest Priority wins; a tie goes to the rule with more criteria, then to the one written first.
What a rule does:
| Key | Effect |
|---|---|
TierCap | the highest tier the first-run benchmark or the GPU heuristic may pick: Low to Cinematic. An explicit r.Quality and a platform's recommended tier are never capped. |
+CVars=name=value | set at device-profile priority: above the tier, below every INI, the player's settings and the command line, so nothing you or the player chose is overwritten |
CVarsUpToTier | the cvars apply only while the resolved tier is at or below this |
Description | shown in the log and in Project Settings > Engine > Scalability |
The shipped profiles
| Profile | Identifies | Tier cap |
|---|---|---|
SteamDeck | Steam Deck LCD and OLED APUs (names, 0x163F, 0x1435, or the platform's word) | Low, plus the cvars below |
NVIDIA_RTX50 | RTX 50 series, 0x2B80-0x2FFF | none |
NVIDIA_RTX40 | RTX 40 series, 0x2680-0x28FF | none |
NVIDIA_RTX30 | RTX 30 series, 0x2200-0x25FF | Epic |
NVIDIA_RTX20 | RTX 20 series and TITAN RTX, 0x1E00-0x1F7F | High |
NVIDIA_GTX16 | GTX 16 series, 0x1F80-0x1FFF, 0x2180-0x21FF | Medium |
AMD_RX9000 | Radeon RX 9000 (RDNA 4), 0x7550-0x75BF | none |
AMD_RX7000 | Radeon RX 7000 (RDNA 3), 0x7440-0x749F | Epic |
AMD_RX6000 | Radeon RX 6000 (RDNA 2), 0x73A0-0x743F | Epic |
AMD_Integrated, AMD_Integrated_RDNA3 | Radeon integrated graphics; the 740M to 890M | Low; Medium |
Intel_ArcB, Intel_ArcA | Arc B-series 0xE200-0xE2FF; Arc A-series 0x5690-0x56BF | Epic; High |
Intel_Integrated, Intel_Integrated_Arc | Intel integrated graphics; integrated Arc (Xe-LPG, Xe2) | Low; Medium |
Apple_M, Apple_M_MaxUltra | Apple M-series; Max and Ultra | High; Epic |
The integrated and Apple profiles also set r.TextureStreaming.PoolSizeMB (1 to 3 GB): their dedicated VRAM is a firmware carve-out, not the memory the GPU reaches.
The Steam Deck profile keeps Low's shadow budget, effects and view-distance cuts and adds Global Illumination from the Medium row (sg.GlobalIlluminationQuality=1), FSR at Quality on the 1280x800 panel (r.Upscaler=3, r.Upscaler.AutoQuality=0, r.UpscalerQuality=1), a 40 fps cap for the panel's 40 Hz mode (t.MaxFPS=40; set 60 for the 60 Hz mode), a dynamic-resolution target of 22 ms, and a GPU-hot alert at 97 °C, because the APU runs near 90 °C under load by design. Its cost has not been measured on a Deck.
What you see
At startup the engine logs the device and the verdict:
[DeviceProfile] gpu='NVIDIA GeForce RTX 3050' vendor 0x10DE device 0x2507 driver 566.36 discrete, 8192 MB VRAM, D3D12, win-x64: NVIDIA_RTX30 (GeForce RTX 30 series (Ampere)): tier cap Epic, 0 cvars
[Quality] preset=High source=auto-tier(... measured GPU index ...)Each cvar the profile sets, or leaves because a higher source holds it, gets its own [DeviceProfile] line. Project Settings > Engine > Scalability shows the matched profile under the tier table.
Test a rule on any machine:
| cvar | Effect |
|---|---|
r.DeviceProfile.Force | use this profile by name, whatever the GPU |
r.DeviceProfile.Enable 0 | no device profile at all |
r.AutoQuality.ForceName, ForceVendorId, ForceDeviceId, ForceVRAMMB, ForceIntegrated, ForceDriver | match as if the GPU were this one |
The driver blocklist
A [DriverBlock <Name>] section names a driver range:
[DriverBlock NVIDIA_Before520]
Vendor=NVIDIA
DriverBelow=520.00
Severity=Warn
Reason=NVIDIA drivers older than 520 cannot run DLSS and miss graphics fixes this game relies on; expect rendering errors and crashes.
Advice=Update to the current GeForce driver from nvidia.com or the NVIDIA app.It takes the identity keys of a profile (Vendor, NameContains, DeviceIds, Platforms, GraphicsApis) and a driver range: MinDriver and MaxDriver are inclusive, DriverBelow exclusive. A section with no driver range would match every driver and is ignored with a warning. Driver versions are the vendor's own: the GeForce number on NVIDIA, the user-mode driver version (31.0.101.5590) on AMD and Intel under D3D, the packed Vulkan version under Vulkan, so a rule for AMD or Intel should name its GraphicsApis.
| Severity | What happens |
|---|---|
Warn | a [DeviceProfile] driver blocklist: ... warning in the log, and in a packaged game a dialog the player sees once per driver version |
Block | an error in the log and, before the first frame, a dialog offering Quit or Continue anyway |
No dialog appears in an unattended run (--unattended, MOWO_NO_CRASH_REPORTER=1) or with r.DriverBlocklist.Prompt 0. The engine ships two entries: NVIDIA drivers older than 520 and Intel Arc drivers older than 31.0.101.4091 (February 2023) on Windows D3D, both Warn.
Streaming to the VRAM you have
The texture streaming pool (r.TextureStreaming.PoolSizeMB, 0 = auto) is sized at startup from the VRAM class of the GPU, leaving room for the render targets (about 1 GB at 1440p with Global Illumination and shadow pages), the meshlet pool, the OS and other processes:
| VRAM class | Pool |
|---|---|
| under 4 GB | 768 MB |
| 4 GB | 1536 MB |
| 6 GB | 2560 MB |
| 8 GB | 3584 MB |
| 12 GB and more | 5120 MB |
Each class starts a little under its nominal size, since a 4 GB card reports less once the driver takes its share. A device profile, an INI or the player's Texture Pool Size setting overrides it.
While the game runs, the VRAM-pressure governor (r.TextureStreaming.VramPressure, on by default) reads the budget the OS grants the process: DXGI QueryVideoMemoryInfo on D3D12 and D3D11, VK_EXT_memory_budget on Vulkan, the recommended working set on Metal. Over 95 % of it, the pool cap drops below what is resident by the overshoot (split between the texture and mesh pools by what each holds), so the farthest textures lose mips first; under 85 % it relaxes and lifts. The log says when it starts and stops capping, and whenever the budget pass changes what it takes away:
[Runtime] Texture streaming: 41 texture(s) 63 mip level(s) below the detail their screen size wants, to fit a 1043 MB pool (VRAM 3010 / 2048 MB used, capped by VRAM pressure)r.TextureStreaming.ForceBudgetMB treats the OS budget as that many megabytes, so the pressure path can be watched on a card with room to spare.
API reference
- SceneRenderer module:
FDeviceProfiles,FDeviceProfile,FDriverBlock,FDeviceProfileInputs,FQualityAutoTier::TexturePoolForVramMB - Game module:
FVramPressureGovernor,FTextureStreamingManager
Engine sources
This page describes behavior implemented in these files of the engine repository, checked against commit 77b5ea1b2.
- Engine/Source/Runtime/SceneRenderer/Public/Renderer/DeviceProfiles.h
- Engine/Source/Runtime/SceneRenderer/Public/Renderer/QualityAutoTier.h
- Engine/Config/BaseDeviceProfiles.ini
- Engine/Source/Runtime/Game/Public/Streaming/VramPressureGovernor.h
- Engine/Source/Runtime/Game/Public/Streaming/TextureStreamingManager.h