XR Daily: A Compiler Ate the Fraction

A Shader Bug That Turned Out to Be a Compiler Bug

Bruno Croci wanted a Voronoi background for a music video, so he wrote the standard Worley noise setup: split space into tiles, scatter a random center in each, then take the distance to the nearest nine centers. It rendered correctly on Linux. On a Windows machine with an RTX 4070, the cells visibly snapped instead of sliding, in every browser he tried.

The most useful part of the writeup is how wrong the first diagnosis was. An LLM blamed Nvidia driver 595.79 and suggested replacing fract(x) with x - floor(x). That substitution removed the stutter, which is exactly the kind of confirmation that ends an investigation early. It was luck. Every attempt to build a minimal repro from the driver theory failed.

The real clue came from changing 398.0 to 398.1 in the noise input. With the decimal intact, the stutter vanished while fract stayed in the source. That sent him into RenderDoc, where a Chromium capture showed the bytecode was ps_5_0, meaning DirectX 11 rather than OpenGL. On Windows, browsers route WebGL through ANGLE, which transpiles GLSL to HLSL for Microsoft’s FXC compiler. Switching Chrome’s ANGLE backend to Vulkan eliminated the bug, and compiling the same HLSL with optimizations disabled did too. The culprit is FXC at O3, dropping a frac it decides cannot see fractional input and truncating coordinates with floor.

The practical lesson concerns the stack under your shader. A Windows user and a Quest user running identical GLSL do not necessarily execute identical instructions, and a bug that appears on one translation path is a portability hazard in your own content, not a driver bug you can report.

→ Bruno Croci

---

Meta Is Reportedly Building Camera-Free Smart Glasses

The Information reports that Meta is preparing a new class of smart glasses with no cameras at all, internally codenamed Luna, targeting a fall launch and an expected appearance at Connect on September 23 and 24 alongside the slim Project Phoenix headset. The described hardware swaps imaging for audio: six onboard microphones for talking to Meta AI and the company’s Muse agent, a side button to trigger the assistant, and speakers in temple arms slimmer than the current frames.

The motivation reads as a response to a privacy fight Meta has been losing in public. New York state and Philadelphia have passed smart glasses legislation, the company faces a class-action suit over footage reportedly sent to a Kenya-based subcontractor for AI training, and its v26 firmware now disables the camera if the capture LED has been physically removed.

Road to VR’s Scott Hayden makes the sharper point: this is not Meta retreating from outward-facing sensors, because AR eventually needs them. It is a second product line for people who will not wear a camera on their face, and it likely costs less to build. For developers the constraint is what ships in the box. Six microphones, a button, and speakers bound you to voice, audio, and head orientation with no visual understanding of the room, which makes this an assistant accessory rather than a spatial computing platform. See also the XR compute offload argument.

→ Road to VR

---

A VR180 Concert That Earns Its Depth

Kacey Musgraves’ Middle of Nowhere arrived on Quest as a 49-minute immersive concert filmed at Billy Bob’s Texas in Fort Worth, built for stereoscopic VR180 rather than adapted from a flat production. Untold Studios produced and creatively directed it, MediaMonks and 7 Cinematics co-produced, and the capture used Blackmagic URSA cameras, with staging, choreography, and lighting designed around a viewer in a headset.

The craft detail is what makes it worth reading. There are not many camera positions, but each one has a job: one sits onstage between the keyboard player and the drummer looking back into the crowd, one stands on the floor looking up toward the stage with the audience around you, and another rides backward as Musgraves and an accordion player walk out onto the dance floor. UploadVR’s Craig Storm watched at the “4K” setting over hotel Wi-Fi and reported natural stereo depth, no eye strain, and spatial audio that did not stand out.

That tradeoff is the honest state of immersive video. VR180 gives you a fixed viewpoint with perfect stereo and no reconstruction artifacts, and it can put you three feet from a performer in a way no 6DOF scene currently matches for photoreal humans. Gaussian splat scenes offer real parallax and freedom to move, but they still struggle with moving subjects and relighting. The production costs sit on opposite sides: choreographed camera work plus a real venue here, capture volume plus compute for splats. See also Gaussian splatting reconstruction.

→ UploadVR

---

Two Environment Breakdowns Worth Reading for 6DOF

80 Level ran two environment art breakdowns on the same day, and together they explain how production artists keep detail density under control. Irmak Çavun broke down an abandoned Middle Eastern courtyard assembled from modular kits and tiling textures. Nore Pollentier walked through a realistic Ankara alley recreated from a real street she located with a reverse image search, then cross-checked on Google Maps for proportions.

Çavun’s workflow change is the transferable part. Instead of a waterfall pipeline, she moved to explicit passes: blockout, then Pass 1 at alpha level with first-pass models and textures, then Pass 2, then polish. Pass 1 only had to be representative, not final, so art direction happened with materials in engine context rather than in a gray Substance viewport. Architecture that resisted the standard 2x4 and 4x4 metre kit sizes became a few core modules plus many smaller snapping pieces, with RGB masking, texture paint, and one modular decal for wall edge damage. The payoff showed up when a courtyard water feature was not reading on screen, and because the maps were barely started she could still save it.

Pollentier’s piece covers the harder case, where the reference is real and therefore never grid-aligned. Her walls were modeled with evenly spaced topology for vertex painting, bricks were sculpted in ZBrush and decimated for Unreal Engine, and pivots were placed so irregular pieces still assemble sanely.

Both pipelines are tuned to a camera the player does not control, which is the part XR developers should borrow carefully. In a headset every wall is reviewed from the floor, from behind, and from inside arm’s reach, so texel density assumptions built for eye-height framing get exposed immediately. Modular kits also keep material count and draw calls down, and that is the real frame budget on a standalone chip. See also environment art breakdowns.

→ 80 Level: Middle Eastern Courtyard · 80 Level: Ankara Street

---

An LLM Labeled 4,290 Comments for $9, Then a 459M Model Took Over

Peter Vijeh scrapes Reddit threads about chef’s knives and extracts brand, model, and steel names. He started by paying Gemini 3.1 Pro per comment, which worked well and scaled badly, because the bill tracked how much strangers posted. The swap he documented is a clean template: have the large model label 4,290 comments once for $9, roughly $0.0021 each, fine-tune GLiNER large v2.5 at 459M parameters on those labels, then run the small model locally for everything after. Zero-shot GLiNER scored about 0.65 F1 against Gemini’s output; after fine-tuning, the best run hit 0.83 F1, with 24 minutes of training on a Tesla T4 and about $2.50 of GPU spend across all ten attempts.

Five of those ten runs produced nothing usable, and none of the failures were about modeling. GLiNER’s default max_steps of 10,000 silently overrode the epoch setting and trained 39 epochs. load_best_model_at_end threw without an eval strategy. The expensive one was a tensor named words_mask, which looks exactly like an attention mask and is not one: it holds incremental word indices, so building it as zeros and ones left the loss flat while everything else looked healthy. The caveat is his own, that nobody hand-checked Gemini’s labels, so the model is graded against Gemini rather than ground truth.

The XR relevance is the shape of the problem. Standalone headsets run under a battery and thermal budget, so anything that executes per frame or per interaction, whether object tagging, intent parsing, or scene semantics, wants a model in the hundreds of millions of parameters rather than a frontier API call. The catch is that the winning run still used a T4-class GPU, so the on-device version means smaller models, harder quantization, and lower expectations than 0.83 F1.

→ Peter Vijeh

---

Nvidia’s 84GB Card Is a Capacity Play, Not a Bandwidth Play

Nvidia’s RTX PRO 5500 Blackwell desktop edition packs 84GB of GDDR7 using 28 modules in a clamshell layout, 14 per side of the PCB, each a 24Gb (3GB) part running at 25 Gb/s over a 448-bit interface. That yields an official 1,398 GB/s of bandwidth and 600W board power. It slots between the RTX PRO 6000 at $15,599 and the RTX PRO 5000 72GB at $9,209, and neither Nvidia nor PNY would give Tom’s Hardware a price, so expect something inside that range.

Tom’s Hardware’s read is that the capacity matters more than the memory speed. Bandwidth lands only about 4% ahead of the RTX PRO 5000 and roughly 22% behind the RTX PRO 6000 and the GeForce RTX 5090, because Nvidia downclocks the GDDR7 it buys at 28 Gb/s to protect higher tiers.

Capacity solves whether something fits: large-scale Gaussian splat training on a full scene, reconstruction with more views, irradiance and lightmap baking at high resolution, avatar codec training. Bandwidth solves how fast it draws, and splat rasterization, per-frame sorting, and foveated rendering passes all want throughput, which is where this card sits behind the 5090 and the 6000. If your pipeline fits a scene offline and ships it to a headset, 84GB is the useful number. If you are chasing real-time performance on a workstation, the bandwidth gap is what shows up in a profiler. See also training a billion Gaussians.

→ Tom’s Hardware

---

Quick Bites

• Great on Frame passes 130 titles. Valve’s compatibility page for Steam Frame verified games and apps now lists more than 130 entries, up from the 70+ announced at launch. → Road to VR
• Into the Radius 2 slips two weeks on PSVR 2. The PlayStation VR2 release of the survival shooter has been pushed back. → UploadVR
• Spymaster leaves early access in October. The espionage adventure exits early access across PC VR and Quest next month. → UploadVR
• Dream Hackers public demo is live. The VR game’s public demo is now available on SteamVR. → UploadVR