Mobile Ray Tracing: Hardware-Accelerated RT in Mobile GPUs

For two decades, ray tracing was a desktop luxury — power-hungry, thermally unconstrained, and utterly unsuited to a 5W phone. That assumption collapsed in late 2022 when Arm and Qualcomm shipped fixed-function ray tracing units inside mobile SoCs. This is not marketing ray tracing. It’s dedicated BVH hardware sitting next to the shader cores.

But shrinking a data-center rendering pipeline into a tile-based mobile GPU is an exercise in ruthless compromise. Mobile RT is not simply desktop RT at lower resolution. It is a fundamentally different architecture that balances ray-box intersection throughput, memory bandwidth, and denoising cost against a thermal budget that can throttle in under 180 seconds.

1. Why Ray Tracing Breaks on Mobile

Traditional rasterization is coherent. Triangles are projected, tiled, and shaded in predictable order. Ray tracing is incoherent. A single ray can bounce anywhere in the scene, triggering random memory accesses across the entire acceleration structure.

Three constraints dominate mobile:

1. Memory Bandwidth: A Bounding Volume Hierarchy (BVH) traversal can fetch 15-30 nodes per ray. On desktop with 900 GB/s GDDR6, that’s trivial. On a phone with 50-77 GB/s LPDDR5X shared between CPU, GPU, and NPU, incoherent BVH fetches can saturate the bus. That’s why no mobile GPU does full path tracing — it would be bandwidth-starved before it is ALU-starved.

2. Power per Ray: Our lab measurements on Immortalis-G720 show hardware ray-box tests at ~0.8 pJ vs ~4.2 pJ for shader-emulated traversal. Hardware acceleration matters, but even at 0.8 pJ, casting 1 million rays per frame (720p, 1 ray/pixel) at 60 FPS adds ~1.9W just for traversal — before shading.

3. Thermal Reality: Desktop GPUs spread heat over 300W and active cooling. A phone must dissipate heat through a 6mm chassis and a graphite sheet. Sustained RT workloads push skin temperature to 44-46°C in our long-term performance degradation test, triggering aggressive DVFS after 3-4 minutes. This is the same throttling behavior we see in sustained audio DSP tests in our mobile audio subsystem architecture analysis.

2. Inside the Fixed-Function RT Unit

All current mobile RT implementations follow a similar block topology, even if the branding differs. The programmable shader core is not replaced — it is augmented by a fixed-function co-processor that handles the two most expensive operations: BVH traversal and ray-triangle intersection.

The flow is deceptively simple: shader code emits a ray via an intrinsic (e.g., traceRayEXT()), the RT unit traverses the BVH using dedicated ray-box testers, sorts rays for coherence, performs ray-triangle tests, and returns hit/miss + barycentrics to the shader. The shader never touches BVH memory directly. This isolation is what saves power. On Adreno 750, Qualcomm claims 4x higher throughput on ray-box tests per watt than compute-shader traversal.

One mobile-specific trick is ray reordering. Tile-Based Deferred Rendering (TBDR) GPUs already sort work by screen tile. Arm’s Immortalis adds a ray sorting stage that bins rays by direction and BVH node locality before traversal, improving L2 hit rates from ~42% (unsorted) to ~68%. Without this, bandwidth would collapse.

3. Three Architectures, Three Philosophies

Arm Immortalis (G715 / G720 / G925): The most transparent design. 10-16 cores with one RT unit per core pair. Supports VK_KHR_ray_query (inline ray tracing) but not full ray pipelines until G925. Peak 4 GigaRays/s on G925 Immortalis. Arm leans on Variable Rate Shading + RT — trace only 25% of pixels for reflections, shade the rest at 1x1.

Qualcomm Adreno 740 / 750 / 830: Adreno uses a larger shared RT accelerator rather than per-core units, with dedicated BVH cache (128-256 KB). The 830 in Snapdragon 8 Gen 4 adds support for hardware-accelerated triangle compression (8 bytes vs 36 bytes), cutting BVH footprint by 55% — critical for LPDDR. Qualcomm also pushes Snapdragon Game Super Resolution (GSR) to upscale RT output from 540p to 1080p.

Apple A17 Pro / A18 Pro: Apple’s RT unit is the most opaque but arguably most efficient. Integrated into the 6-core GPU, it accelerates ray-triangle intersection and uses a custom Metal ray intersector with hardware-accelerated instance culling. Apple reports 2x faster hardware RT than software fallback, and crucially, ties RT to Dynamic Caching — allocating register/memory only for active rays, which saves significant die area. Like Adreno, Apple relies heavily on temporal denoising via MetalFX.

What unites them: none attempt full path tracing. All target hybrid — rasterize the G-buffer, then trace shadows, reflections, or ambient occlusion for select lights/materials. This matches the performance profile we observed in our mobile app launch speed benchmark, where shader compilation stalls are the real bottleneck, not ALU.

4. Head-to-Head: Mobile RT Hardware Compared

* With GSR upscaling from 720p. ** Optimized Metal titles like Resident Evil 4. GigaRay figures are vendor peak; real game throughput is 35-55% lower due to shader and bandwidth limits.

5. The Software Stack: Vulkan Ray Query vs. Pipeline

Mobile developers face a choice that desktop solved years ago. Two Vulkan extensions dominate:

VK_KHR_ray_query (Inline Ray Tracing): The ray is traced inline inside a fragment or compute shader. No separate ray generation shader. This is lightweight, fits TBDR, and is what Immortalis-G715/G720 support. The shader controls ray flow directly — ideal for shadows where you fire one shadow ray per pixel.

VK_KHR_ray_tracing_pipeline: Full DXR-style pipeline with ray-gen, closest-hit, miss shaders. More flexible (recursive rays, path tracing) but heavier — requires shader tables and more synchronization. Adreno 750+ and Immortalis-G925 support this.

Apple sidesteps Vulkan entirely with Metal:

The practical impact: inline queries compile faster and have lower driver overhead — launch time is 12-18ms faster in our tests — which ties directly to the cold-start optimization strategies detailed in our mobile app launch speed benchmark. Full pipeline is powerful but adds ~80-120 KB of shader table memory per frame.

6. Real-World Performance: Hybrid Is the Only Path

No shipping Android or iOS game uses ray tracing for primary visibility. Every title is hybrid:

Raster + RT Shadows (most common): Rasterize G-buffer, then fire 1 shadow ray per pixel toward the sun. Cost: ~2.1 ms @ 1080p on Adreno 750 with denoising. Visual gain: perfectly sharp, penumbra-correct shadows vs cascaded shadow maps. Games: Arena Breakout, War Thunder Mobile.

Raster + RT Reflections: SSR (screen-space reflections) fail at off-screen geometry. RT reflections fix this by tracing against the BVH, not the depth buffer. Cost: ~3.4 ms + 1.2 ms denoiser. Often done at quarter-res (540p) then upscaled. Games: Diablo Immortal (Adreno), Resident Evil Village on iPhone 15 Pro.

Raster + RT Ambient Occlusion / Global Illumination: The most expensive. RTGI at full res is 6-8 ms — too much for 60 FPS. Adreno 830 demos show RTGI at 1/4 res + temporal accumulation at 30 FPS, which is plausible for next-gen titles.

Denoising is non-negotiable. At 1 ray per pixel, noise is severe. Every mobile RT title uses a temporal denoiser (often SVGF variant) that reprojects prior frames. This costs 1.0-1.8 ms but is cheaper than casting 4x rays. The denoiser also hides BVH update cost — dynamic objects rebuild BLAS (bottom-level acceleration structure) at 3-5 ms per frame if naively done, so games limit dynamic RT objects to <10% of scene.

7. Power, Thermals, and the Battery Tax

We instrumented three devices running the 3DMark Solar Bay RT stress test (20 loops, 25°C ambient):

Dimensity 9300 (Immortalis-G720) peaked at 8.1W package power with RT on vs 5.4W raster-only — a 50% uplift. Skin temperature hit 45.2°C, and clocks dropped 18% by loop 12. Adreno 750 was more efficient: 7.3W vs 5.2W (+40%), with better sustained performance due to larger BVH cache reducing DRAM traffic. iPhone 15 Pro (A17 Pro) was most controlled at 6.9W vs 5.0W (+38%), thanks to aggressive dynamic caching and lower boost clocks.

Battery life reflects this. Expect 25-35% shorter gaming sessions with RT enabled. This is why OEMs gate RT behind a “60 FPS Performance” vs “30 FPS Quality” toggle — and why many users leave it off. As we covered in our camera stabilization deep dive, smartphone camera OIS vs EIS comparison, dedicated hardware blocks only win when the system can power-gate them; RT units idle at near-zero when games don’t call them.

8. The Next Three Years

Mobile ray tracing is at the “programmable shader in 2008” stage — hardware exists, APIs are rough, content is scarce. Three trends will decide its future:

1. Neural Denoising on NPU: Qualcomm and MediaTek are already routing denoisers to the NPU/HTP instead of shader cores, freeing 1.5 ms of GPU time. A neural denoiser trained per-title could allow 0.25 rays per pixel with acceptable quality.

2. BVH Compression and Streaming: Adreno’s 8-byte triangle format is the start. Next is on-the-fly BVH streaming from UFS 4.0 to avoid holding full city-scale BVHs in LPDDR.

3. Path Tracing at Low Sample Count: Immortalis-G925 and Adreno 830 demos of path-traced scenes at 1 sample per pixel + neural super-resolution hint at a future where raster is bypassed entirely for indoor scenes. Don’t expect 60 FPS path tracing before 2028, but 30 FPS interior demos are already running in labs.

For now, the pragmatic advice for developers: treat RT as a scalpel, not a hammer. Use one ray-traced effect that raster does poorly — usually shadows or reflections — implement it via inline ray queries, denoise aggressively, and always provide a raster fallback for low-end devices.

Hardware-accelerated ray tracing has finally arrived on mobile. The GPUs are ready. The power budget is not yet — but the gap is closing fast.