One test game map produced 56 TB of raw wave simulation data, and the encoder stored it in 41 MB, a compression factor of over a million (Raghuvanshi and Snyder, ACM Transactions on Graphics 2014). At SIGGRAPH 2026, a Ubisoft La Forge and Audiokinetic team cut one uncompressed parameter family from 182 GB to 14 MB, and 28 game audio professionals rated the result 61.8 against 62.6 for the reference (Seuté et al., SIGGRAPH Conference Papers 2026). Real-time ray tracing runs on a far tighter budget: Valve’s Steam Audio plugin defaults to 4,096 rays and 4 bounces (Valve, Steam Audio Unity plugin source). We aggregated data from ACM Transactions on Graphics, ACM SIGGRAPH, IEEE Virtual Reality, the Journal of the Acoustical Society of America, and vendor documentation from Valve, AMD, NVIDIA and Meta. Where our spatial audio statistics track devices and consumer demand, this roundup tracks the machinery underneath: compute cost, data size and what listeners actually hear.
TL;DR
- 56 TB of raw simulation data encoded to 41 MB for one test map (Raghuvanshi and Snyder, ACM TOG 2014).
- 182 GB of uncompressed wave coding data versus 14 MB for a latent model, rated 61.8 versus 62.6 for realism (Seuté et al., SIGGRAPH 2026).
- Early reflections became distinguishable once mean free path, the average distance sound travels between reflections, changed by more than 0.06 m (Rungta et al., P-Reverb, 2019).
- Steam Audio defaults to 4,096 real-time rays and 4 bounces (Valve, Steam Audio Unity plugin source).
- A GPU compute framework held 80-100 FPS at 16,000 rays per frame on an RTX 2060 (University of Gothenburg thesis).
- A hybrid ray-traced and neural pipeline took 6.78 to 12.95 ms per frame across four scenes (Tang et al., IEEE VR 2021).
- A neural generator produced an acoustic effect in 0.1 milliseconds on an RTX 2080 Ti (Ratnarajah and Manocha, IEEE VR 2024).
- 10th-order diffraction ran up to 568 times faster than the previous state of the art (Schissler et al., SIGGRAPH 2021).
- The gap between simulated and measured impulse responses was always clearly audible across nine test environments (Brinkmann et al., JASA 2019).
- Reverberation method moved perceived distance in a dynamic scene, F(1,16) = 15.29, p < 0.01 (Rungta et al., 2017).
1. Wave Bakes: Terabytes In, Megabytes Out
Wave simulation solves the physics of sound directly, so occlusion (a blocked sound going muffled) and diffraction (sound bending around an edge) emerge from the math rather than from a hand-tuned filter. It is far too slow for a running game, so studios bake it: they compute the acoustics offline and ship only the results. A factor of over a million separates what the bake computes from what the player installs.
That split decides who can afford the technique. A bake measured in hours on a 140-machine cluster becomes a very different proposition for a small team: one developer’s report works out to about 778 hours on a single 8-core machine if run back to back (3,500 probes × 800 seconds ÷ 3,600). Each probe, a sample point in the level where sound is simulated, can be sent to a different compute node, which is why the same thread points to cloud baking. The approach is not a lab curiosity: Microsoft Research lists Gears of War, Sea of Thieves and Borderlands 3 among titles built on it.
| Metric | Value | Source |
|---|---|---|
| Raw simulation data vs encoded runtime data, Sanctuary test map | 56 TB vs 41 MB | Most recent available data: Raghuvanshi and Snyder, ACM TOG 2014 |
| Bake time and encoded size across the paper’s five test maps | 4 to 20 hours on a 140-machine cluster; 20.8 MB to 53.7 MB | Most recent available data: Raghuvanshi and Snyder, ACM TOG 2014 |
| Developer-reported bake load on one project | 3,500 probes at 800 seconds per probe on an 8-core CPU | Most recent available data: GitHub, microsoft/ProjectAcoustics issue 55, May 27, 2021 (one developer’s report) |
| Uncompressed data for one acoustic parameter family, wave coding vs latent model | 182 GB vs 14 MB (Wwise Audio Lab scene); 3.1 GB vs 1.8 MB (Audio Gym scene) | Seuté et al., Reciprocal Latent Fields, SIGGRAPH Conference Papers 2026 |
| Boundary element simulation to build a neural training set | About 12 days for about 100,000 objects, up to 1,000 Hz | Most recent available data: Tang et al., IEEE VR 2021 |
| Boundary element simulation time per object, 125 Hz vs 1,000 Hz | 0.87 s vs 2.80 s | Most recent available data: Tang et al., IEEE VR 2021 |
| Bake time for a ray-traced acoustic map, not a wave simulation (Meta’s own claim) | Seconds to minutes, depending on scene complexity | Meta, Acoustic Ray Tracing for Unity documentation |
Comparison note: 41 MB is a compressed figure, while 182 GB and 14 MB are uncompressed, so the two rows are not directly comparable. The 2026 paper treats compression as out of scope and reports no ratio. Our math on its largest scene is 182,000 MB ÷ 14 MB = 13,000 to 1.
2. Runtime Cost: Microseconds per Source
Baking moves the pain out of the frame. Once the data exists, looking it up is close to free, and that is the whole argument for the technique: about 100 microseconds per source on an Intel Xeon E5540 at 2.53 GHz, with roughly 100 active sounds supported. The cost the player’s machine pays is a small lookup, and the expensive part already happened on someone else’s cluster.
Neural models now compete in the same range. The newest reports sit in the microseconds to low milliseconds, though they are not measured on equal terms: some use a single CPU thread, others a high-end GPU, and a GPU that is spending its frame on graphics is not free. Read this table as orders of magnitude, not as a ranking.
| Metric | Value | Source |
|---|---|---|
| Decode time per source, averaged over 100,000 random point pairs without cache | About 100 μs; about 100 active sounds | Most recent available data: Raghuvanshi and Snyder, ACM TOG 2014 |
| Acoustics query time on PC, version 2.0 vs 3.0 (Microsoft’s own claim) | Cache hit (data already in memory) 25 to 15 μs; cache miss 1,100 to 40 μs | Most recent available data: Microsoft, Project Acoustics 3.0 announcement, August 31, 2022 |
| Latent model inference per 100 source-receiver pairs, one CPU thread | 10 μs to 93 μs for the three latent variants; 247 μs for the large comparison network | Seuté et al., SIGGRAPH 2026 |
| Time to generate one acoustic effect for a source and listener position | 0.1 milliseconds on an RTX 2080 Ti | Ratnarajah and Manocha, Listen2Scene, IEEE VR 2024 |
| Neural impulse responses generated for a furnished indoor scene | More than 10,000 per second on an RTX 2080 Ti | Most recent available data: Ratnarajah et al., MESH2IR, ACM Multimedia 2022 |
| Frame time, ray tracing plus neural scattering prediction, four scenes | 6.78 ms to 12.95 ms; prediction itself ≤1 ms per view | Most recent available data: Tang et al., IEEE VR 2021 |
| VRWorks Audio filters per source and listener pair (NVIDIA’s own claim) | 2 filters, up to 48,000 Hz and up to 2 seconds long | NVIDIA, VRWorks Audio SDK in Depth |
An impulse response is a recording of how a space answers a single click. Applying it to a dry sound is called convolution, and the number of filters that step needs grows quickly with spatial detail (see section 4).
3. Ray Budgets: Thousands, Not Millions
Audio ray tracing lives on a small allowance, because every ray competes with the rest of the frame for time. A default of 4,096 rays is the number that most clearly marks the scale of the real-time problem. Steam Audio traces those rays from the listener, so the setting describes one budget for the scene rather than a figure per source.
Shipped games rarely publish this number. Massive Entertainment says Avatar: Frontiers of Pandora sends out “thousands of rays” every frame on the GPU for sound obstruction, inside a hybrid system (the studio’s own description, with no exact count). Obstruction means a partial block where sound still reaches the listener another way, in contrast to full occlusion. Research systems report far deeper tracing, 50 to 200 reflection orders in the table below against a default of 4 bounces in Steam Audio. The gap is the distance between what a lab demo can spend and what a default setting assumes.
| Metric | Value | Source |
|---|---|---|
| Default rays and bounces, real-time vs baked (Valve’s default setting) | 4,096 rays and 4 bounces vs 16,384 rays and 16 bounces | Valve, Steam Audio Unity plugin source, master branch |
| Default real-time limits (Valve’s default setting) | 32 sources (128 maximum); 5% of CPU cores | Valve, Steam Audio Unity plugin source |
| Ceiling on occlusion samples per source (Valve’s default setting) | 16 by default, range 1 to 128 | Valve, Steam Audio Unity plugin source |
| GPU compute shader ray tracing in a complex architectural scene, RTX 2060 | 16,000 rays per frame at around 80-100 FPS | University of Gothenburg thesis, A GPU Compute Framework for Real-Time Ray-Traced Acoustic Simulation in Unity |
| Reflection orders traced for late reverberation | 200 orders across seven octave bands, 125 Hz to 8,000 Hz | Most recent available data: Tang et al., IEEE VR 2021 |
| Reflection orders supported at interactive rates on a multi-core PC | 50 to 100 orders of specular and diffuse reflections | Most recent available data: Rungta et al., 2017 |
| Average simulation time per sample point, three scenes, early vs late reflections | 16.7 ms to 47.5 ms vs 68.4 ms to 150.7 ms | Most recent available data: Rungta et al., P-Reverb, 2019 |
Those P-Reverb timings are per point during precomputation, not per frame during play. The reverberation engineers measure in a game is the same physical quantity that drives our concert hall reverberation statistics, only computed under a time limit of milliseconds.
4. Speedups: Claimed vs Measured
Every headline multiplier in this field needs its denominator checked. The 50-150X figure for GPU baking compares against one vendor’s own built-in tracer, for baking reverb, on two scenes of 80,000 and 140,000 triangles. That is a useful fact about offline tools, and it says nothing about what happens to a live frame.
GPU compute also has a physical cap. Doubling the compute units reserved for audio gives less than double the speed, and audio has to share the chip with graphics. The other limit is convolution: moving from zero-order to third-order Ambisonics, a format that resolves sound from sixteen directions, multiplies the filter count by 16. Two of the independent papers below report gains of 3x to 5x, and those come from doing less work (fewer rays, fewer sample points) rather than from running the same work on faster silicon.
| Metric | Value | Source |
|---|---|---|
| Reverb baking speedup vs Steam Audio’s built-in ray tracer (Valve’s own claim) | 50-150X with Radeon Rays on OpenCL 1.2 GPUs; around 20X with Embree on recent 4-core CPUs | Most recent available data: Valve, Steam Audio news, October 30, 2019 |
| Speedup when doubling GPU compute units for Radeon Rays, 4 to 8 or 8 to 16 (Valve’s own claim) | About 1.2x to 1.9x | Most recent available data: Valve, Steam Audio news, February 21, 2019 |
| Convolution filters for 16 sources, third-order vs zero-order Ambisonics (AMD’s own demo) | 256 vs 16 | Most recent available data: AMD GPUOpen, TrueAudio Next acceleration of Steam Audio, October 21, 2019 |
| Computation time from adaptive ray tracing with an impulse response cache, vs previous approaches | 5x speedup | Most recent available data: Schissler and Manocha, i3D 2016 |
| Memory overhead of the same method, vs previous approaches | About two orders of magnitude lower | Most recent available data: Schissler and Manocha, i3D 2016 |
| Speedup from clustering sample points by perceptual thresholds, three scenes | 3x to 4x | Most recent available data: Rungta et al., P-Reverb, 2019 |
| Neural impulse response generation vs a geometric acoustic algorithm on a CPU | More than 200 times faster | Most recent available data: Ratnarajah et al., MESH2IR, ACM Multimedia 2022 |
Outlier note: the MESH2IR multiplier compares a GPU network with a CPU algorithm, so part of the gap is hardware rather than method. Valve’s per-scene values appear only in a chart image, so we report the ranges stated in its text.
5. Occlusion and Accuracy: Where the Errors Live
The number closest to occlusion itself is direct sound level, which one 2026 paper defines as the parameter capturing distance attenuation and obstruction. Even the best-performing latent variant there misses it by 1.9 dB in a small scene and 4.2 dB in a larger one, and the 2014 pipeline reported a similar scale of loudness error. These are errors against a reference simulation, not against a real room.
Against real rooms the picture is less flattering. A round robin in the Journal of the Acoustical Society of America found the difference from measurement was always clearly audible, and the algorithms were mostly plausible rather than authentic. The authors link that gap largely to simplified absorption and scattering data and to weak modeling of diffraction, the exact effect occlusion depends on. Measured references of this kind are why anechoic chamber testing still matters for simulation work.
| Metric | Value | Source |
|---|---|---|
| Mean error of the full encoded wave pipeline | 3.3 dB for loudness; 22.5% for decay rate | Most recent available data: Raghuvanshi and Snyder, ACM TOG 2014 |
| Mean absolute error of the best latent variant, Audio Gym scene vs Wwise Audio Lab scene | Direct sound level 1.9 dB vs 4.2 dB; direction of arrival 3.8 vs 3.2 degrees | Seuté et al., SIGGRAPH 2026 |
| Normalized reproduction error of predicted acoustic scattering, overall and by frequency | 8.8% overall; 4.2% at 125 Hz rising to 10% at 1,000 Hz | Most recent available data: Tang et al., IEEE VR 2021 |
| Deviation of predicted acoustic metrics from ground truth | Under 10% | Most recent available data: Ratnarajah et al., MESH2IR, ACM Multimedia 2022 |
| Geometric estimate of mean free path (average distance sound travels between reflections) and reverberation time | Within 3% and within 4.6% | Most recent available data: Rungta et al., P-Reverb, 2019 |
| Speed of finding 10th-order diffraction paths vs the previous state of the art | Up to 568 times faster | Most recent available data: Schissler, Mückl and Calamia, SIGGRAPH 2021 |
| Room simulation round robin: simulated vs measured impulse responses, six acoustic scenes and three complex environments | Difference “always clearly audible”; “mostly plausible but not authentic” | Most recent available data: Brinkmann et al., JASA 2019 |
Ray budgets show up here too. The Gothenburg thesis found its GPU tracer modeled early reflections and direction well but could not sustain late reverberation: energy decayed faster than the wave reference because of ray budget limits and compute shader divergence. The thesis concludes that GPU audio ray tracing should be used primarily as a secondary tool that improves perceptual quality and spatial accuracy.
6. Perception: What Listeners Actually Notice
The most useful result in this field may be how often accuracy stops mattering. Game audio professionals rated a 14 MB learned model as realistic as the full reference, p = 0.41, while a simple free-space model scored less than half as high. In that test, geometry-aware acoustics was clearly audible, yet the remaining physical error was not, which suggests structure may matter more than the last few decibels.
Context also matters. Distance judgments changed with the reverberation method in a dynamic scene but not in a static one, and a matched room beat a highly reverberant one, though not an anechoic one, in a small 2026 presence study. Generic listening hardware adds its own limits on top, as our gaming headset HRTF localization data shows, so simulated acoustics are only as good as the last stage of the chain.
| Metric | Value | Source |
|---|---|---|
| Mean realism rating (0-100), latent model vs wave coding reference vs free-space low anchor, 28 game audio professionals | 61.8 ± 3.8 vs 62.6 ± 3.9 vs 27.7 ± 4.1 | Seuté et al., SIGGRAPH 2026 |
| Statistical comparison of the same ratings | Latent model vs reference p = 0.41; both vs low anchor p < 0.001, Cohen’s d ≈ 1.4 | Seuté et al., SIGGRAPH 2026 |
| Effect of reverberation method on perceived distance, dynamic scene, 17 participants | F(1,16) = 15.29, p < 0.01 | Most recent available data: Rungta et al., 2017 |
| Power-law fit of perceived distance (Dr) to real distance (dr), ray tracing vs Schroeder filter | Dr = 1.56·dr^0.58 (R² = 0.94) vs Dr = 1.08·dr^0.66 (R² = 0.99) | Most recent available data: Rungta et al., 2017 |
| Effect of reverberation method on perceived distance, static scene, 12 subjects | No significant effect | Most recent available data: Rungta et al., 2017 |
| Scenes with a significant difference in realism ratings across three rendering pipelines, 51 qualified online participants of 71 recruited | 2 of 4: Floor χ² = 10.82, p < 0.01; Havana χ² = 8.27, p = 0.02 | Most recent available data: Tang et al., IEEE VR 2021 |
| Spatial presence in Unreal Engine 5 on Meta Quest 3, ten critical listeners | Significant drop only between matched (RT60 = 0.21 s) and highly reverberant (RT60 = 2.0 s) rooms | Spacek, Lee and Johnson, AES AVAR Conference 2026 |
| Smallest noticeable change in mean free path for early reflections, 106 online participants | 0.06 m | Most recent available data: Rungta et al., P-Reverb, 2019 |
Context note: the 2026 realism test used pre-recorded videos with a fixed source position and receiver trajectory, so raters did not control the camera. The 51-participant study ran online on Amazon Mechanical Turk. Neither design tests competitive play.
Summary: Ray-Traced Audio Occlusion and Acoustic Simulation in Games by the Numbers
| Metric | Value | Source |
|---|---|---|
| Raw simulation data vs encoded data, one test map | 56 TB vs 41 MB | Raghuvanshi and Snyder, ACM TOG 2014 |
| Decode time per source | About 100 μs | Raghuvanshi and Snyder, ACM TOG 2014 |
| Wave coding vs latent model, uncompressed, one parameter family | 182 GB vs 14 MB | Seuté et al., SIGGRAPH 2026 |
| Realism rating, latent model vs reference, 28 professionals | 61.8 vs 62.6, p = 0.41 | Seuté et al., SIGGRAPH 2026 |
| Steam Audio default real-time rays and bounces | 4,096 and 4 | Valve, Steam Audio Unity plugin source |
| Rays per frame at around 80-100 FPS, RTX 2060 | 16,000 | University of Gothenburg thesis |
| Frame time, ray tracing plus neural scattering, four scenes | 6.78 ms to 12.95 ms | Tang et al., IEEE VR 2021 |
| Normalized reproduction error of predicted scattering | 8.8% | Tang et al., IEEE VR 2021 |
| Time to generate one acoustic effect, RTX 2080 Ti | 0.1 milliseconds | Ratnarajah and Manocha, IEEE VR 2024 |
| Neural impulse responses per second, RTX 2080 Ti | More than 10,000 | Ratnarajah et al., ACM Multimedia 2022 |
| 10th-order diffraction vs previous state of the art | Up to 568 times faster | Schissler, Mückl and Calamia, SIGGRAPH 2021 |
| Adaptive ray tracing speedup vs previous approaches | 5x | Schissler and Manocha, i3D 2016 |
| Convolution filters, 16 sources, third-order Ambisonics (AMD’s own demo) | 256 | AMD GPUOpen, October 21, 2019 |
| Reverberation method effect on perceived distance, dynamic scene | F(1,16) = 15.29, p < 0.01 | Rungta et al., 2017 |
| Smallest noticeable change in mean free path, early reflections | 0.06 m | Rungta et al., P-Reverb, 2019 |
| Spatial presence, matched vs highly reverberant room, ten listeners | Significant drop | Spacek, Lee and Johnson, AES AVAR 2026 |
| Simulated vs measured impulse responses, nine test environments | Always clearly audible | Brinkmann et al., JASA 2019 |
Methodology and Sources
- Raghuvanshi and Snyder, Parametric Wave Field Coding for Precomputed Sound Propagation, ACM Transactions on Graphics 33(4), 2014
- Seuté, Vasudev, Richan and Buffoni, Reciprocal Latent Fields for Precomputed Sound Propagation, SIGGRAPH Conference Papers 2026
- Tang, Meng and Manocha, Learning Acoustic Scattering Fields for Dynamic Interactive Sound Propagation, IEEE VR 2021
- Ratnarajah and Manocha, Listen2Scene, IEEE VR 2024
- Ratnarajah, Tang, Aralikatti and Manocha, MESH2IR, ACM Multimedia 2022
- Rungta, Rewkowski, Klatzky, Lin and Manocha, Effects of Virtual Acoustics on Dynamic Auditory Distance Perception, 2017
- Rungta, Rewkowski, Klatzky and Manocha, P-Reverb, 2019
- Schissler and Manocha, Adaptive Impulse Response Modeling for Interactive Sound Propagation, i3D 2016
- Schissler, Mückl and Calamia, Fast Diffraction Pathfinding for Dynamic Sound Propagation, SIGGRAPH 2021
- Brinkmann, Aspöck, Ackermann, Lepa, Vorländer and Weinzierl, A Round Robin on Room Acoustical Simulation and Auralization, JASA 2019
- Spacek, Lee and Johnson, The Role of Room Acoustics in 6DoF Interactive Audio for XR, AES AVAR Conference 2026 program
- University of Gothenburg, A GPU Compute Framework for Real-Time Ray-Traced Acoustic Simulation in Unity
- Valve, Steam Audio Unity plugin settings source and Steam Audio news
- AMD GPUOpen, TrueAudio Next Acceleration of Steam Audio Sound Reflections
- NVIDIA, VRWorks Audio SDK in Depth
- Meta, Acoustic Maps in Acoustic Ray Tracing for Unity
- Project Acoustics 3.0 announcement and Project Triton overview
- Massive Entertainment, Ray Tracing Audio in Snowdrop
- GitHub, microsoft/ProjectAcoustics issue 55
- Data watch: much of this field’s best public data is old, so rows dated before October 2023 are flagged in their tables. Shipped games almost never publish ray counts or audio frame budgets, and we found no such figure for any title. Vendor figures are labeled as claims, and no market-size estimate is cited because we found none specific to acoustic simulation. The Tang et al. paper states network overhead both as under 11 ms per object and under 1 ms per view; we report the per-view figure from its results table. The thesis page we read showed no author or year, and the AES AVAR entry is a program abstract, not a full paper. Many academic rows share one research circle around Dinesh Manocha, so they are distinct studies but not fully independent voices. The bake load from GitHub is a single developer’s report from the 2.0 era, and we could not confirm that the VRWorks Audio SDK is still maintained or that Microsoft’s Project Acoustics documentation remains current, since two of its pages redirected to the general Xbox developer hub when we checked in October 2026.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the ACM SIGGRAPH Asia 2026 technical papers are published.