ARM Mali GPU Architecture

Sam Martin
Graphics Architect, ARM

ARM Game Developer Day - London

03/12/2015
 Agenda
 Mali architecture and tiling introduction

 Behind the scenes – power limits

 Vulkan

2 © ARM 2015
 Mali GPU Taxonomy In a Nutshell
 Mali 4xx series OpenGL ES 2.0
 1-8 shaders cores, separate fragment and vertex processors

 Mali 6xx – 8xx OpenGL ES 3.x

 Unified “tri-pipe” shader core
 Larger core configurations, max 16 cores from Mali 760 +
 AFBC, ASTC, Transaction Elimination, ...

 All tile-based GPUs

3 © ARM 2015
 Command stream Command phase
from CPU

Input assembly
Geometry phase
Vertex shader

Rasterizer

Pixel shader Pixel phase

Output merger

4 © ARM 2015
 Tile-based GPUs
Command stream
Input assembly
from CPU
 Fragments >> Geometry
Vertex shader

Rasterizer

Pixel shader
 Phased structure
1. Buffer all operations into “render passes” Outer merger
2. Transform + bin all geometry into screen space tiles
3. Fully shade each tile into local memory, then write back

5 © ARM 2015
 Mali Architecture
 Hardware tiling

 Forward Pixel Kill

 Reduce overdraw

 Framebuffer memory on-chip

 4x MSAA for “free”
 Advanced on-chip shading

 Bandwidth efficiencies
 ARM Framebuffer Compression
 Transaction elimination
 ASTC

6 © ARM 2015
 Mobile Power Limits
 Lifetime constrained by battery Phones 1-3 Watts
 High-end performance constrained by heat
Tablets 3-5 Watts

 Thermal Design Power/Point (TDP) Small laptop-like 10-25 Watts

 Capacity constrained by ability to dissipate heat
Regular laptop 25-50 Watts
 Memory bandwidth particularly expensive Integrated desktop 40-100 Watts
 Rule of thumb: 100mW / GB/s, assume 1 W total

 Low-mid end GPUs are constrained by die area

 Savings prolong battery life but may not increase performance

7 © ARM 2015
3 mm² 5 mm² 10 mm² 30 mm²

561 mm²

Similarly capable mobile GPUs NVIDIA GeForce

Die areas shown to scale GTX Titan
8 © ARM 2015
3 mm² 5 mm² 10 mm² 30 mm²

561 mm²

Low-end

9 © ARM 2015
3 mm² 5 mm² 10 mm² 30 mm²

561 mm²

Mid-range

10 © ARM 2015
3 mm² 5 mm² 10 mm² 30 mm²

561 mm²

High-end

11 © ARM 2015
3 mm² 5 mm² 10 mm² 30 mm²

561 mm²
 1-10x range, just within mobile phones
 Servicing such a wide range demands scalable GPU designs
 GPU feature set cannot indicate performance capability

12 © ARM 2015
 Thermal Throttling
 CPU - big
 CPU - LITTLE
 GPU
GL Benchmark 2.7 (T-Rex HD) [3 Runs]

Max OPP big

Max OPP LITTLE

Frequency

Max OPP GPU

Median filtered chart for clarity

Time (s)
13 © ARM 2015
 Thermal Throttling
 CPU - big
 CPU - LITTLE
 GPU
GL Benchmark 2.7 (T-Rex HD) [3 Runs]

Max OPP big

Max OPP LITTLE

Frequency

Max OPP GPU

Median filtered chart for clarity

Time (s)
14 © ARM 2015
 Thermal Throttling
 CPU - big
 CPU - LITTLE
 GPU
GL Benchmark 2.7 (T-Rex HD) [3 Runs]

Max OPP big

Max OPP LITTLE

Frequency

Max OPP GPU

Median filtered chart for clarity

Time (s)
15 © ARM 2015
 Thermal Throttling
 CPU - big
 CPU - LITTLE
 GPU
GL Benchmark 2.7 (T-Rex HD) [3 Runs]

Max OPP big

Max OPP LITTLE

Frequency

Max OPP GPU

Median filtered chart for clarity

Time (s)
16 © ARM 2015
 Thermal Throttling
 CPU - big
 CPU - LITTLE
 GPU
GL Benchmark 2.7 (T-Rex HD) [3 Runs]

Max OPP big

Max OPP LITTLE

Frequency

Max OPP GPU

Median filtered chart for clarity

Time (s)
17 © ARM 2015
 Vulkan
 Good match for mobile and tiling architectures
 Explicit multi-pass render passes
 No hidden costs (copies, allocs, shader recompiles, etc)
 Multi-threaded
 Low overhead

 Gloves-off API
 Needs care – look out for future info post-release

18 © ARM 2015
 Thanks! Questions?

[email protected] @palgorithm

 Coming up:
 Increase texturing efficiency and quality
 Daniele Di Donato, “Get the most out of ASTC” – up next!
 Advanced use of tiled framebuffers
 Marius Bjørge, “Fast Approximate Indirect Lighting on Mobile”, 11am
 Compute shaders & tessellation
 Hans-Kristian Arntzen, “Real-time GPU-driven Ocean Rendering on Mobile”, 11.30am

19 © ARM 2015
 For more information visit the Mali
Developer Centre:

http://malideveloper.arm.com

• Revisit this talk in PDF and audio

format post event
• Download tools and resources

20 © ARM 2015
The trademarks featured in this presentation are registered and/or unregistered trademarks of ARM Limited (or its
subsidiaries) in the EU and/or elsewhere. All rights reserved. All other marks featured may be trademarks of their
respective owners.
Copyright © 2015 ARM Limited