DLSS 4.5 on RX 9000 via d4r 0.1.2: Native FP8, No Card Tested [2026]
By Indie Kings | October 1, 2026
Updated October 1, 2026: Native FP8 DLSS 4 and 4.5 kernels for RX 9000 land in d4r 0.1.2 per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. No physical RX 9000 card has been tested yet so there are no published performance or image quality results for RDNA 4 hardware.
Image: AMD header art. Credit: AMD.
What d4r 0.1.2 Adds
d4r 0.1.2 adds native DLSS 4 and DLSS 4.5 kernels for gfx1201 and gfx1200 builds per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. The gfx1201 target covers the RX 9070 XT, the RX 9070, the GRE variant, and the AI PRO R9700. The gfx1200 builds cover the RX 9060 XT and the RX 9060. This is a build level addition with kernel code paths aimed at RDNA 4, not a claim of tested speed or tested image quality on those cards.
The release notes frame NativeFp8 as the default path on RDNA 4 per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. A 16 bit fallback remains available in a manner similar to RDNA 3. That detail matters because precision paths affect compatibility across GPU generations. Native FP8 is positioned as the RDNA 4 route, while the fallback keeps older behavior reachable when FP8 is not appropriate or not available. Readers should treat this as a technical capability added to the code, not as proof of a frame rate gain.
Scope remains narrow by design. d4r 0.1.2 supports Super Resolution only per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. There is no Frame Generation support in this path. There is no Ray Reconstruction support in this path. There is no Neural Rendering support in this path. Anyone looking for a full DLSS feature stack on Radeon through this tool will not find it in the stated 0.1.2 scope.
The supported runtime is DX12 on Linux and Proton per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. That limits the audience to Linux gamers, Steam Deck style handheld researchers, dual boot testers, and Proton curious Windows users willing to test outside Windows. It is not presented as a Windows native drop in replacement. It is not presented as a universal DX11 plus DX12 plus Vulkan solution. The DX12 plus Linux plus Proton boundary should shape expectations before anyone downloads anything.
The delivery method runs the official nvngx_dlss.dll via ZLUDA plus ROCm plus custom kernels per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. That stack explains why version tracking matters so much here. An official NVIDIA DLL supplies the DLSS implementation, ZLUDA supplies translation compatibility, ROCm supplies AMD compute support, and custom kernels supply the hardware specific glue. Change one layer and behavior can shift. This is why point releases such as 0.1.1 to 0.1.2 deserve careful reading rather than quick hype.
For search clarity, the headline claim is narrow: DLSS 4.5 on RX 9000 via d4r 0.1.2 means kernel support exists in the build, not that RX 9000 performance has been measured. That distinction will recur through this article because early social posts tend to blur it. Kernel support means the code has a path for those GPU IDs. Tested support would mean a physical card produced repeatable frame rates and screenshots. The first is verified as added. The second has not happened in public per the sourced status update. Keeping those ideas apart protects readers from buying hardware on a misunderstanding.
Version context also helps. d4r is an early stage compatibility effort, and 0.1.2 follows 0.1.1 with incremental work rather than a finished product launch. Small version numbers signal active development, changing behavior, and incomplete validation. Gamers used to polished driver releases should reset expectations. A 0.1.x tool can be exciting for testers while remaining unsuitable for a primary gaming PC. Back up settings, expect troubleshooting, and do not treat early builds as stable software [NEEDS VERIFICATION].
The RX 9000 naming spread adds confusion worth clearing. gfx1201 is the larger RDNA 4 target in this release, while gfx1200 is the smaller companion build target. The article maps gfx1201 to RX 9070 XT, RX 9070, GRE, and AI PRO R9700, and gfx1200 to RX 9060 XT and RX 9060, per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. Those mappings come from the sourced report, not from hands on PCI ID testing by this site. Anyone validating support should check the d4r repository device notes directly before assuming a specific SKU is covered.
Finally, prior local coverage handles DLSS 5 on AMD via a different mod path. That separate path is out of scope here and this article does not link it because slugs were unverified this session. Readers should not mix DLSS 5 mod claims with d4r 0.1.2 claims. They involve different code, different authors, and different compatibility limits [NEEDS VERIFICATION]. The clean way to think about it is one article per path, one version per test, one GPU per result. Anything looser invites false comparisons.
How It Works
d4r works by letting AMD hardware execute the official nvngx_dlss.dll through a compatibility stack built on ZLUDA plus ROCm plus custom kernels per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. In plain terms, the game still asks for DLSS Super Resolution as if NVIDIA hardware were present. The translation layer intercepts that request and reroutes execution to AMD friendly kernels. The official DLL remains the source of DLSS logic rather than a clean room reimplementation. That design choice preserves behavior while shifting execution to Radeon.
ZLUDA handles the CUDA compatibility problem at a high level. Many DLSS paths assume NVIDIA CUDA behavior, so unmodified execution on AMD would fail. ZLUDA translates those expectations into something ROCm and Radeon can process [NEEDS VERIFICATION]. ROCm then provides the AMD compute platform underneath. Custom kernels finish the job by implementing precision and architecture specific routines for each supported gfx target. Each layer has a distinct role, which helps diagnose failures when a game crashes or produces corrupt output.
NativeFp8 as default on RDNA 4 is the headline technical change in 0.1.2 per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. FP8 refers to 8 bit floating point math, which can reduce memory traffic and increase throughput when the hardware and model support it. RDNA 4 includes improved AI throughput relative to RDNA 3, which makes an FP8 path plausible as a default [NEEDS VERIFICATION]. The 16 bit fallback mirrors RDNA 3 behavior and preserves a safer route when FP8 is unstable or unsupported. Precision defaults are important, but they do not by themselves guarantee higher frame rates.
Super Resolution only scope keeps the problem tractable. Super Resolution upscales a lower resolution render to a higher output resolution using temporal data and trained models. Frame Generation would need to synthesize whole intermediate frames. Ray Reconstruction would need to replace denoisers in ray traced pipelines. Neural Rendering would imply broader model driven rendering features. None of those are included here per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. Testers should therefore compare d4r against other upscalers, not against a full DLSS 3 or DLSS 4 frame generation stack.
The Linux and Proton focus reflects where this stack is testable today. DX12 on Linux and Proton is the stated runtime per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. Proton translates Windows game calls for Linux, which creates a convenient insertion point for DLL level work. A native Windows path would face different loader behavior, driver signing issues, and anti cheat scrutiny [NEEDS VERIFICATION]. Linux testers accept those tradeoffs more readily. That is one reason early results cluster around Proton friendly single player titles rather than competitive multiplayer games.
Installation flow in general terms involves placing the correct DLSS DLL, installing the d4r build matched to the GPU architecture, and configuring Proton or DX12 settings so the game loads the intended path [NEEDS VERIFICATION]. Exact file names, environment variables, and launch options change between releases, so this article does not provide step by step commands. Readers should follow the official d4r GitHub instructions relayed via VideoCardz rather than copying random Reddit comments. Version mismatch is the most common source of black screens and fallback to native resolution rendering.
Image quality evaluation needs care. DLSS version numbers such as DLSS 3 versus DLSS 4 versus DLSS 4.5 refer to evolving models and presets, not just speed. A higher version can look cleaner at the same Quality preset, or it can cost more performance for extra detail. On AMD through translation, model behavior may differ from NVIDIA native behavior [NEEDS VERIFICATION]. Screenshots need matched settings, matched scenes, and lossless capture to mean anything. Frame counters alone cannot show ghosting, shimmer, or temporal instability.
Performance evaluation needs equal care. Upscaler tests must lock resolution, preset, and scene before comparing builds. The 7700 XT data discussed later uses 1440p Quality in one game, which is a reasonable template. Changing drivers, changing Proton versions, or changing in game settings between runs invalidates comparison. Early stage tools also show run to run variance, so multiple runs matter more than a single peak number. Readers should view any single FPS tweet as a hint, not as a benchmark.
Security and stability notes deserve space even without a tutorial. Loading unofficial translation layers plus official DLLs into games can trigger anti cheat systems [NEEDS VERIFICATION]. Single player Proton testing is lower risk than multiplayer testing, but no third party graphics mod is risk free. Save games can be affected when a renderer crashes mid write in rare cases, so backups are prudent. None of this means d4r is unsafe by design. It means early graphics research belongs on a test setup rather than on a ranked account or a work machine.
Emulation vs Hardware Status
Status is the most misunderstood part of this story. gfx1201 is checked through emulation only, and gfx1200 is compiled only, per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. No physical RX 9000 performance results have been published. No physical RX 9000 image quality results have been published. That is the full public status for RDNA 4 hardware in this release.
Emulation only means the code path was exercised without a physical RX 9070 XT or sibling card producing gameplay numbers. Emulation can catch build errors, interface mismatches, and obvious crashes. It cannot prove frame rates, frame pacing, power behavior, or visual stability in a shipping game. Compiled only is even narrower. It means the gfx1200 target builds successfully, not that it was functionally validated in gameplay. Both states are normal for early bring up, but neither equals tested support.
The support table below summarizes the public state in one place. It separates build availability from hardware validation. Readers scanning quickly should focus on the status column rather than the target column. A listed target without tested status is a development milestone, not a buying signal.
| Target | Status |
|---|---|
| gfx1201 for RX 9070 XT, RX 9070, GRE, AI PRO R9700 | Emulation only, no physical card tested, no performance or image quality results published |
| gfx1200 builds for RX 9060 XT, RX 9060 | Compiled only, no physical card tested, no performance or image quality results published |
| RDNA 3 reference example RX 7700 XT | Tested in Silent Hill Townfall at 1440p Quality with published FPS across 0.1.1 to 0.1.2 |
| Super Resolution on DX12 Linux and Proton | Supported scope in d4r 0.1.2 |
| Frame Generation, Ray Reconstruction, Neural Rendering | Not supported in this path |
Why does emulation come first. Hardware availability, price, and developer access often lag code readiness. A maintainer may add RDNA 4 kernel paths before owning every SKU. Community testers may own the cards but lack time to run controlled tests. Emulation lets development proceed without blocking on hardware loans. The tradeoff is uncertainty. Until someone publishes a physical RX 9000 run, emulation status should be read as progress, not proof [NEEDS VERIFICATION].
What would change the status. A credible RX 9000 report needs GPU model, driver version, distro plus kernel, Proton version, d4r version, DLSS DLL version, game title plus patch, resolution plus preset, and a repeatable scene. Ideally it includes minimum frame rates and frame time data alongside averages, plus screenshots or video with matched settings. One tweet with a single average FPS number would be a start, but it would not settle image quality. Three independent testers with matched methods would be far stronger than one anonymous screenshot.
What readers should not do now. Do not buy an RX 9000 card expecting DLSS 4.5 gains based on 0.1.2 build notes. Do not assume emulation success predicts equal performance to NVIDIA native DLSS. Do not assume compiled gfx1200 code will run flawlessly on launch day. Hardware buys should rest on tested FSR 4 performance, raster performance, price, power, and features available today [NEEDS VERIFICATION]. Treat d4r RDNA 4 support as a research lead to watch, not as a purchase justification.
What watchers should track next. The next meaningful update would be a physical RX 9070 XT run with versioned methods. After that, an RX 9060 XT compiled build moving from compiled only to emulated to physically tested would show maturation. Then cross game testing beyond a single title would show generality. Finally, preset scaling from Performance through Quality through Balanced would show whether FP8 behavior holds across modes [NEEDS VERIFICATION]. Each step is separate. Skipping straight to broad claims would repeat the hype cycle this article aims to avoid.
Attribution discipline matters here. The emulation only and compiled only facts come per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. Any Reddit thread claiming physical RX 9000 numbers after that report needs fresh verification before citation. This article deliberately avoids embedding unverified Reddit performance claims. If new numbers appear, they should be checked for DLL version, Proton version, and scene repeatability before being treated as real.
The 7700 XT Numbers That Exist
The only concrete numbers in the sourced report come from an RX 7700 XT, not from RX 9000 hardware. The test case is Silent Hill Townfall at 1440p Quality per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. Across d4r 0.1.1 to 0.1.2, DLSS 4.5 ranges from 48.4 to 49.3 FPS, DLSS 4 ranges from 67.9 to 68.1 FPS, and DLSS 3 ranges from 71.2 to 71.9 FPS. Those are the published reference points. Everything else about RDNA 4 speed remains unmeasured in public.
The table below restates those figures without adding interpretation in the cells. Keep the game, resolution, and preset fixed when reading. Cross version comparisons inside one GPU are valid context. Cross GPU comparisons to RX 9000 would be invalid because no RX 9000 row exists yet.
| DLSS Version via d4r on RX 7700 XT | Game and Settings | FPS Range across d4r 0.1.1 to 0.1.2 |
|---|---|---|
| DLSS 4.5 | Silent Hill Townfall 1440p Quality | 48.4 to 49.3 FPS |
| DLSS 4 | Silent Hill Townfall 1440p Quality | 67.9 to 68.1 FPS |
| DLSS 3 | Silent Hill Townfall 1440p Quality | 71.2 to 71.9 FPS |
The pattern is clear even before analysis: newer DLSS model versions cost more performance on this AMD path. DLSS 3 is fastest in raw FPS, DLSS 4 is slower, and DLSS 4.5 is slowest by a wide margin on the 7700 XT sample. That ordering does not prove DLSS 4.5 looks worse. It suggests DLSS 4.5 does more work per frame or hits a less optimized path through translation. Image quality could still be higher at lower FPS, but no pixel level comparison is provided in the sourced numbers. Speed and quality must be judged together, and only speed is published here.
Below FSR 4 is the key context per VideoCardz. The sourced report states performance sits below FSR 4 on this hardware, and that FSR 4 still makes more sense for most gamers per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. That guidance is practical rather than tribal. FSR 4 is the vendor native upscaler path on supported Radeon hardware, with driver and game integration that a translation layer cannot match for ease of use [NEEDS VERIFICATION]. A research tool can be slower and harder to run while remaining interesting to developers. Most players should choose the faster supported option.
Single game limits deserve emphasis. Silent Hill Townfall is one title with one engine, one lighting mix, and one scene selection. Upscaler behavior varies with motion, foliage, particles, and ray tracing load. A result in a dark corridor horror scene may not transfer to an open world racing game or a fast shooter [NEEDS VERIFICATION]. One resolution plus one preset is also narrow. Performance scaling at 1080p Performance versus 4K Quality can differ sharply because bandwidth and compute bottlenecks shift. Treat the 7700 XT figures as existence proof that d4r runs, not as a general performance rating.
Version to version movement looks small. The ranges from 0.1.1 to 0.1.2 span less than one FPS for DLSS 4 and DLSS 3, and under one FPS for DLSS 4.5. That stability is mildly encouraging for reproducibility. It does not show a breakthrough optimization in 0.1.2 for this GPU. The RDNA 4 relevant change in 0.1.2 is the new FP8 kernel path, which would not be expected to transform RDNA 3 results. Readers should therefore avoid crediting 0.1.2 with a 7700 XT speed win. The numbers show continuity, not a leap.
What is missing for a fair verdict. No minimum FPS or percentile data is quoted. No frame time charts are quoted. No power draw is quoted. No image comparison sliders are quoted. No blind test votes are quoted. Those gaps are normal for an early status relay, but they prevent strong conclusions about playability. A 49 FPS average can feel different with smooth pacing versus stutter, and screenshots can reveal whether extra cost buys visible clarity. Until those artifacts arrive, the numbers answer whether it runs, not whether it wins.
How to use these numbers if you own a 7700 XT. First, keep FSR 4 as the default unless you enjoy testing [NEEDS VERIFICATION]. Second, if you test d4r, match the template: same 1440p Quality preset, same game patch, same Proton build, same DLL, then record multiple runs. Third, capture screenshots in matched scenes for DLSS 3 versus 4 versus 4.5 so you can judge whether lower FPS buys cleaner edges. Fourth, share methods alongside results so others can reproduce them. That workflow turns curiosity into useful data instead of noise.
How not to use these numbers. Do not extrapolate 7700 XT FPS to a 9070 XT. RDNA 4 has different compute throughput, memory behavior, and FP8 support, so scaling is not linear [NEEDS VERIFICATION]. Do not assume DLSS 4.5 will remain slowest on RDNA 4 native FP8. The new kernel path could shift the ordering once physical cards are tested. Do not present 48.4 FPS as a universal DLSS 4.5 on AMD rating. It is one GPU plus one game plus one preset plus one stack version. Precision in claims keeps community testing honest.
Should You Bother
For most gamers, the sourced guidance stands: FSR 4 still makes more sense per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. That sentence should anchor buying and settings decisions. Native vendor support usually means simpler setup, broader game coverage, and performance tuned for the hardware in the machine [NEEDS VERIFICATION]. d4r is a compatibility achievement worth following, but it is not a reason to change GPU brands or to delay a build. Play today with what is tested, watch d4r as research.
Linux and Proton testers are the exception. If you already run DX12 games under Proton, enjoy version pinning, and can document methods, d4r 0.1.2 is worth a careful look [NEEDS VERIFICATION]. The Super Resolution only scope keeps testing focused. The 7700 XT template shows what to record. The RDNA 4 kernel additions give testers with 9070 or 9060 class hardware something specific to validate. Contribution quality matters more than enthusiasm here. Clean logs help more than hype posts.
Windows only gamers should wait. The stated runtime is DX12 on Linux and Proton per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. Forcing an early Linux stack onto a Windows only setup adds friction without a proven payoff. There is also no published RX 9000 result to chase. Waiting costs nothing because FSR 4 remains available. Early adoption makes sense when the upside is measured. Here the RDNA 4 upside is still unmeasured.
Competitive multiplayer gamers should skip this build entirely. Third party DLL translation can interact poorly with anti cheat, and any renderer instability carries heavier consequences in ranked play [NEEDS VERIFICATION]. Single player Proton titles with save backups are the appropriate venue. Even there, test one game at a time and keep a known good configuration to revert. Stability habits matter more than preset bravado when running 0.1.x graphics code.
Content creators and benchmarkers should treat 0.1.2 as a methods story, not as a ranking story. The interesting video is not DLSS beats FSR. The interesting video documents setup, shows version strings on screen, runs matched scenes across DLSS 3 versus 4 versus 4.5, and zooms into motion artifacts [NEEDS VERIFICATION]. Viewers learn more from transparent methods than from a thumbnail FPS number. Given the lack of RX 9000 hardware results, any creator claiming RDNA 4 leadership from this release would be overstating.
Budget planning deserves realism. An RX 9060 XT or RX 9060 purchase should be justified by raster performance, VRAM, power, price, and FSR 4 game support today [NEEDS VERIFICATION]. An RX 9070 XT or RX 9070 purchase should rest on the same tested factors plus higher tier raster and feature needs. d4r FP8 support is a free option ticket at best. It might mature into a useful alternative upscaler, or it might remain a niche tool. Do not pay extra today for an untested tomorrow.
Risk checklist before testing includes backups, version notes, and rollback plans [NEEDS VERIFICATION]. Record distro, kernel, Mesa or AMD driver, Proton version, d4r version, DLSS DLL version, game patch, resolution, and preset. Test a repeatable scene rather than wandering. Capture averages plus minimums when possible. Share failures as well as successes because negative results prevent wasted effort. That discipline is what turns emulation only status into physically tested status over time.
Long term outlook stays open but unproven. Native FP8 kernels for RDNA 4 could narrow the gap versus native paths once real cards are measured [NEEDS VERIFICATION]. Broader game coverage could follow if Super Resolution proves stable across engines. Frame Generation and Ray Reconstruction remain out of scope, so even success would leave gaps versus full stack DLSS. The honest summary is progress without proof. Follow the repository, wait for physical card data, and keep FSR 4 as the practical default until measurements say otherwise.
FAQ
Does d4r 0.1.2 support DLSS 4.5 on RX 9000?
It adds native DLSS 4 and 4.5 kernels for gfx1201 and gfx1200 builds per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit, but no physical RX 9000 card has been tested and no performance or image quality results are published.
Which RX 9000 cards are targeted?
gfx1201 covers RX 9070 XT, RX 9070, GRE, and AI PRO R9700 while gfx1200 builds cover RX 9060 XT and RX 9060 per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit.
What does it run?
It runs the official nvngx_dlss.dll via ZLUDA plus ROCm plus custom kernels on DX12 Linux and Proton with Super Resolution only per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit.
What are the real FPS numbers?
On RX 7700 XT in Silent Hill Townfall at 1440p Quality, DLSS 4.5 is 48.4 to 49.3 FPS, DLSS 4 is 67.9 to 68.1 FPS, and DLSS 3 is 71.2 to 71.9 FPS across 0.1.1 to 0.1.2, all below FSR 4 per VideoCardz.
Does it support Frame Generation or Ray Reconstruction?
No. This path supports Super Resolution only with no Frame Generation, Ray Reconstruction, or Neural Rendering support per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit.
Should most gamers use it instead of FSR 4?
No. FSR 4 still makes more sense for most gamers per VideoCardz, so treat d4r as Linux and Proton research until physical RX 9000 results appear [NEEDS VERIFICATION].
Bottom Line
d4r 0.1.2 is a real step for RDNA 4 kernel coverage with NativeFp8 default and 16 bit fallback, but it remains unproven on physical RX 9000 hardware. gfx1201 is emulation only and gfx1200 is compiled only, with no published FPS or image comparisons per VideoCardz by WhyCry on October 1 2026 relaying Countervolts d4r GitHub and Reddit. The only hard numbers come from RX 7700 XT testing where newer DLSS versions run slower and trail FSR 4. Most gamers should stay on FSR 4 and let Linux testers produce the first reproducible RX 9000 runs.
Sources: VideoCardz report by WhyCry October 1 2026 and Countervolts d4r GitHub repository via VideoCardz. Unverified context in this draft is marked [NEEDS VERIFICATION].