ARM Holdings PLC · NASDAQ:ARM
ARM — Technical Presentation (2026)
29 slides from ARM's technical presentation, Aug 7, 2026
Slide 1: State of GPU Driven Rendering on Mobile
The Premier Conference & Exhibition on Computer Graphics & Interactive Techniques State of GPU Driven Rendering on Mobile
Slide 2: About me
About me RODRIGO HOLZTRATTNER Dev Tech Lead at Snapdragon Game Studios, Qualcomm, working directly with game studios on CPU/GPU architecture and optimization https://www.linkedin.com/in/rodrigoholztrattner/
Slide 3: GPU-Driven Spectrum
GPU-Driven Spectrum • The industry often discusses GPU-driven rendering as a single destination, but in practice is an evolving feature. • Mobile renderers gradually shift responsibilities from the CPU to the GPU. • The question is not whether an engine is GPU- driven, but how GPU-driven it is. Leve
Slide 4: GPU-Driven Spectrum
GPU-Driven Spectrum • Mobile GPU-driven must respect tilers, bandwidth, thermal limits, sync, and API portability. It is not the desktop problem adopted downward. • The pressure is already here: Nanite, Lumen, clustered lighting, GPU particles, mesh shaders. • Content scale is growing faster than CP
Slide 5: GPU-Driven Spectrum
GPU-Driven Spectrum • Mobile GPU-driven must respect tilers, bandwidth, thermal limits, sync, and API portability. It is not the desktop problem adopted downward. • The pressure is already here: Nanite, Lumen, clustered lighting, GPU particles, mesh shaders. • Content scale is growing faster than CP
Slide 6: GPU-Driven Spectrum
GPU-Driven Spectrum Technique Mobile status Notes Nanite (or equivalent)* Shipped Available on high-end devices Lumen (or equivalent)* Shipped UE5.4+ partial support on high-end Android deferred path MegaLights (or equivalent)* Prototype / Shipping First few mobile games coming soon Mesh / Task / Ob
Slide 7: Mobile Landscape
Mobile Landscape % of Top-100 Revenue High-fidelity 3D Native / Game Ports 12% ~27% potential candidates Advanced 3D 15% Simple 3D / Stylized 15% Casual 2D / Idle / Puzzle 58% -5 5 15 25 35 45 55 65 Based on published renderer analyses, developer conference disclosures, and store category research.
Slide 8: Mobile Landscape
Mobile Landscape Adoption of demanding rendering techniques takes time − Deferred rendering was rare on mobile a decade ago. Today it is the default for high-fidelity 3D. Indirect draw and mesh shaders are following the same curve. The 27% share today is the starting point, not the ceiling. © 2026 S
Slide 9: Mobile Landscape
Mobile Landscape Adoption of demanding rendering techniques takes time − Deferred rendering was rare on mobile a decade ago. Today it is the default for high-fidelity 3D. Indirect draw and mesh shaders are following the same curve. The 27% share today is the starting point, not the ceiling. Experime
Slide 10: Experiment
Experiment Simple Ball Collision/Occlusion Simulation • Stress benchmark for specific simulation scenarios − Tuned for GPU driven sim • Multi-LOD • Multi-material (32 different ball material groups) • Fullscreen lighting/shadow pass • Tilt sensor because it feels right Tested on Snapdragon 8 Elite G
Slide 11: Experiment: Full CPU
Experiment Full CPU (don’t do this) • Simulation + Collision: CPU • Culling: CPU Frustum • Batched draw • Obviously inefficient FPS Power GPU Usage GPU Freq 14.6 FPS 5.59 W 82.3% 815 MHz I wasn’t aware our lab devices could still be operable at such high temperatures
Slide 12: Experiment: GPU Indexed
Experiment GPU Indexed (best in slot) • Simulation + Collision: GPU • Culling: CPU Frustum • Batched draw • Standard/traditional rendering • Stable FPS Power GPU Usage GPU Freq 121.4 FPS 2.02 W 82.8% 656 MHz
Slide 13: Experiment: GPU Indirect
Experiment GPU Indirect • Simulation + Collision: GPU • Culling: GPU Frustum + GPU HZB • Batched draw • Thermal hit at ~4.3 mins mark FPS Power GPU Usage GPU Freq 84.4 FPS 3.84 W 98.7% 1044 MHz
Slide 14: Experiment: GPU Indirect
Experiment GPU Indirect • Simulation + Collision: GPU • Culling: GPU Frustum + GPU HZB • Batched draw • Thermal hit at ~4.3 mins mark Total Mem Write Text Read Cycles Text Instructions FPS Power GPU Usage GPU Freq Indexed 1215 16 2 12.1 84.4 FPS 3.84 W 98.7% 1044 MHz Indirect 1479 42 10 40.5 Adreno
Slide 15: Experiment: GPU Mesh Shaders
Experiment GPU Mesh Shaders • Simulation + Collision: GPU • Culling: GPU Frustum + GPU HZB • Batched draw • Thermal hit at ~4.1 mins mark FPS Power GPU Usage GPU Freq 75.1 FPS 3.77 W 98.7% 1012 MHz
Slide 16: Why is traditional rendering this competitive?
Why is traditional rendering this competitive? • GPUs (mobile or not) are known to predict which CPU Draw[0], Draw[1], Draw[2] … Draw[N] resources will be needed according to what is being submitted • Commands without GPU indirection clearly defines Command Buffer what and when things will run SetPi
Slide 17: Why is traditional rendering this competitive?
Why is traditional rendering this competitive? • GPUs (mobile or not) are known to predict which CPU ComputeDispatch → DrawIndirect(buffer) resources will be needed according to what is being submitted • Commands without GPU indirection clearly defines Indirect Buffer what and when things will run ?
Slide 18: Why is traditional rendering this competitive?
Why is traditional rendering this competitive? • We are slowly moving optimizations from submission (CPU) level into execution level (GPU) • GPU can perform certain optimizations dynamically according to the current or previous frame info Compute Dispatch Early Z Pass DrawIndirect Generates draw arg
Slide 19: New GPU features follow a maturation pattern
New GPU features follow a maturation pattern • Hardware ray tracing launched on mobile (2022–2023) with significant power cost Power Consumption • By 2024: second-generation RT hardware on high-end mobile 8W 7W delivers roughly 2× the RT throughput. 6W ~89% improvement • The same pattern applies to
Slide 20: GPU-driven today
GPU-driven today What mobile GPU-driven games normally share Game A Game B Game C Game D Game E • Base algorithm: HZB current last last (2x) history + last rast − PreZ − HZB Cull instance instance instance cluster+inst inst+sphere − HZB test MVP 2.7 MB 30 MB 60 MB 7 MB (FP16) 1 MB − Visibility − ICB
Slide 21: GPU-driven today
GPU-driven today What mobile GPU-driven games normally share Game A Game B Game C Game D Game E • Base algorithm: HZB current last last (2x) history + last rast − PreZ − HZB Cull instance instance instance cluster+inst inst+sphere − HZB test MVP 2.7 MB 30 MB 60 MB 7 MB (FP16) 1 MB − Visibility − ICB
Slide 22: GPU-driven today: How can YOU do it today
GPU-driven today How can YOU do it today • The biggest challenges are: − Memory management/bandwidth − Occupancy − Work scheduling • The good news: The same principles that make Lumen, Nanite, and modern GPU-driven renderers efficient can be applied to any engine today.
Slide 23: GPU-driven today: Experiment often
GPU-driven today: Experiment often • Tile shading / tile shaders • Device-specific optimizations (texture x buffer balance, 16-bit vs 32-bit data and arithmetic) • Persistent tile memory: bind buffers to the tile heap so they survive pass boundaries • Break large compute passes into smaller ones − e
Slide 24: GPU-driven today: Shader optimizations
GPU-driven today: Shader optimizations What to do Watch out for • For 16-bit image loads, use relaxed precision • Scalarization is not always a win. Many mobile GPUs (RelaxedPrecision decoration via inline SPIR-V in have scalar registers (uGPRs) but the uniform register HLSL). file is limited. • Use
Slide 25: GPU-driven today: Shader optimizations
GPU-driven today: Shader optimizations What to do Watch out for • For 16-bit image loads, use relaxed precision • Scalarization is not always a win. Many mobile GPUs (RelaxedPrecision decoration via inline SPIR-V in have scalar registers (uGPRs) but the uniform register HLSL). file is limited. • Use
Slide 26: GPU-driven today: Indirect draw generation
GPU-driven today: Indirect draw generation Generate and consume pitfall What to do • Compute shader generates draw commands, then • Move the draw-generating compute shader earlier in the DrawIndirect is issued immediately after. frame. • Insert other GPU work between the compute dispatch and • Drive
Slide 27: GPU-driven today: Other
GPU-driven today: Other • Respect storage buffer size limitations − Be mindful of the meshlet sizes and their impact • High geometry scenes benefit hugely from GPU driven rendering • Meshlet culling is fast and effective • Depth buffer culling of meshlets is great, if you have depth • Bindless resou
Slide 28: Experiment: After Improvements
Experiment: After Improvements GPU Indirect GPU Mesh Shaders New New New New FPS Power GPU Usage GPU Freq FPS Power GPU Usage GPU Freq Old 84.4 FPS 3.84 W 98.7% 1044 MHz Old 75.1 FPS 3.77 W 98.7% 1012 MHz New 107.2 FPS 3.43 W 98.4% 1026 MHz New 88.6 FPS 3.58 W 98.0% 1008 MHz
Slide 29: Thank you
The Premier Conference & Exhibition on Computer Graphics & Interactive Techniques Thank you