Recommended graphics settings per GPU: rework the tiers (an RTX 5090 gets less than Ultra) #965

New issue

Open

opened 2026-10-23 07:15:17 +00:00 by sky_line · 2 comments

sky_line commented

2026-10-23 07:15:17 +00:00

Owner

Problem

On a fresh settings file, an RTX 5090 (Vulkan GPU tier class 5: discrete, hardware RT, 31 GB VRAM) gets recommended settings that land somewhere between High and Ultra -- applyRecommendedSettings applies level 5 from the feature tables (indra/newview/featuretable*.txt), which still carry LL's GL-era class / level mapping. The top cards should come up at the top tier, and every card class should get the settings that actually suit it under our Vulkan renderer.

Wanted

Decide the tiers together (owner + dev): for each GPU class (by capability: VRAM, hardware RT, vendor / architecture, iGPU vs discrete, Pascal / Turing / Ampere / Ada / Blackwell, RDNA 2/3/4, Intel Arc / iGPU, translation layers), which quality level and which individual settings (RT shadows, RT reflections, GI / DI, probes, shadow resolution, AO tier, draw distance, texture budget, upscaler + render scale, weather tier, async compute, texture compression, VRS) are the default.
Implement the mapping from measured capabilities (we already classify in vkClassifyDevice / initVulkanForRealUI: class, RT, VRAM) instead of LL's GL feature-table levels; keep it data-driven (one table we can tune), log the chosen tier and why.
Re-run the recommendation on a GPU change (and offer it when the user asks), never silently overwrite settings a user has changed.
Use the known reference points: RTX 5090 / 4090 = Ultra (+ RT), mid-range (3060 / 5800X datapoint), Pascal 1080 targets (FS-equivalent ~15 ms at 4K, high settings 30+ FPS), 8 GB cards (texture / RT VRAM starvation), RTX 3050 / 4 GB (#946 small weather tier).

Related: #824 (deferred lighting default per GPU class), #405 (stated minimum / one budget), #946 (small cards), #483 (show when RT is inactive and why).

Test

Fresh settings on each reference card: the logged tier and the applied settings match the table.
A user-changed setting survives a re-detection.

## Problem On a fresh settings file, an RTX 5090 (Vulkan GPU tier class 5: discrete, hardware RT, 31 GB VRAM) gets recommended settings that land somewhere between High and Ultra -- `applyRecommendedSettings` applies level 5 from the feature tables (`indra/newview/featuretable*.txt`), which still carry LL's GL-era class / level mapping. The top cards should come up at the top tier, and every card class should get the settings that actually suit it under our Vulkan renderer. ## Wanted 1. **Decide the tiers together (owner + dev):** for each GPU class (by capability: VRAM, hardware RT, vendor / architecture, iGPU vs discrete, Pascal / Turing / Ampere / Ada / Blackwell, RDNA 2/3/4, Intel Arc / iGPU, translation layers), which quality level and which individual settings (RT shadows, RT reflections, GI / DI, probes, shadow resolution, AO tier, draw distance, texture budget, upscaler + render scale, weather tier, async compute, texture compression, VRS) are the default. 2. **Implement the mapping** from measured capabilities (we already classify in `vkClassifyDevice` / `initVulkanForRealUI`: class, RT, VRAM) instead of LL's GL feature-table levels; keep it data-driven (one table we can tune), log the chosen tier and why. 3. Re-run the recommendation on a GPU change (and offer it when the user asks), never silently overwrite settings a user has changed. 4. Use the known reference points: RTX 5090 / 4090 = Ultra (+ RT), mid-range (3060 / 5800X datapoint), Pascal 1080 targets (FS-equivalent ~15 ms at 4K, high settings 30+ FPS), 8 GB cards (texture / RT VRAM starvation), RTX 3050 / 4 GB (#946 small weather tier). Related: #824 (deferred lighting default per GPU class), #405 (stated minimum / one budget), #946 (small cards), #483 (show when RT is inactive and why). ## Test - Fresh settings on each reference card: the logged tier and the applied settings match the table. - A user-changed setting survives a re-detection.

sky_line added the

feature

label

2026-10-23 07:15:17 +00:00

sky_line added this to the Slipstream project

2026-10-23 07:15:17 +00:00

sky_line commented

2026-10-23 07:16:29 +00:00

Author

Owner

Owner direction: on top-end cards we should launch with aggressive defaults that show exactly how good Slipstream can look, chosen from settings we know still perform well there (measured, not guessed). Note: the RC's quality slider showed 6 only because it was moved by hand; the recommendation itself applied level 5 (between High and Ultra) on an RTX 5090.

So for the top class(es): Ultra-plus defaults -- the RT features, GI / DI, high-quality probes, shadows and draw distance the owner's 5090 runs at a good frame rate -- with an upscaler mode where it buys the headroom, verified against the GPU timers on reference scenes (club, dense sim, snow / weather).

Owner direction: on top-end cards we should launch with **aggressive** defaults that show exactly how good Slipstream can look, chosen from settings we **know** still perform well there (measured, not guessed). Note: the RC's quality slider showed 6 only because it was moved by hand; the recommendation itself applied level 5 (between High and Ultra) on an RTX 5090. So for the top class(es): Ultra-plus defaults -- the RT features, GI / DI, high-quality probes, shadows and draw distance the owner's 5090 runs at a good frame rate -- with an upscaler mode where it buys the headroom, verified against the GPU timers on reference scenes (club, dense sim, snow / weather).

sky_line commented

2026-10-23 07:17:57 +00:00

Author

Owner

Reference machines for tuning the tiers:

Top: RTX 5090 (owner) -- the showcase / Ultra-plus table.
Low-mid: RTX 3060 (a tester's machine).
Low: GTX 1080 test box (Pascal; no hardware RT) -- the low-end table and the Pascal targets.
Plan: set the top look on the 5090 and measure frame times on the reference scenes, then derive and verify the lower tiers on the 3060 and the 1080.

Reference machines for tuning the tiers: - **Top:** RTX 5090 (owner) -- the showcase / Ultra-plus table. - **Low-mid:** RTX 3060 (a tester's machine). - **Low:** GTX 1080 test box (Pascal; no hardware RT) -- the low-end table and the Pascal targets. Plan: set the top look on the 5090 and measure frame times on the reference scenes, then derive and verify the lower tiers on the 3060 and the 1080.

sky_line self-assigned this

2026-10-26 02:58:23 +00:00

sky_line referenced this issue

2026-10-26 05:43:14 +00:00

Intel Arc (Mesa): ray-traced shadows 3.5-15 FPS in crowded regions vs 25-55 with shadow maps #1104