ARM Holdings PLC · NASDAQ:ARM
ARM — Qualcomm Technologies — Technical Presentation (2015)
16 slides from ARM's technical presentation, 2015
Slide 1: Xroads of Discovery SIGGRAPH 2015
Xroads of Discovery BO SIGGRAPH 2015 2
Slide 2: Mobile HW and Bandwidth
Mobile HW and Bandwidth Your logo on white Andrew Gruber Qualcomm Technologies, Inc.
Slide 3: Agenda and Goals
Agenda and Goals • Describe the Power and Bandwidth challenges facing Mobile Graphics • Describe some of the Power Saving approaches used in Mobile Graphics to deal with these challenges – By no means a systematic account, just a sample of techniques used by major providers • Provide hints and best
Slide 4: LPDDRx Bandwidth Evolution
LPDDRx Bandwidth Evolution LPDDR4 PC-DDR4 BW LPDDR3
Slide 5: Mobile Power Consumption for a Game
Mobile Power Consumption for a Game • Game running at 1080p / 60 fps • DDR consumption is ~16% for this gaming use case on Qualcomm ® Snapdragon™ processors. – Comparable to internal GPU power and CPU power consumption Mobile device - Power Breakdown Typical OpenGL ES 3.1 based game Rest of CPU the
Slide 6: Resolutions
Resolutions • Mobile Resolutions are high and increasing 1280x720 Base Resolution 1920x1080 2.25x more pixels 2560x1440 4x more pixels 3840x2160 9x more pixels
Slide 7: Binning/Tiling Architecture to Save Memory Bandwidth
Binning/Tiling Architecture to Save Memory Bandwidth GPU has a dedicated fast tile buffer. Rendering Order The rendering surface is split up into “bins”. Rendering pass draws only pixels that are visible (saving GPU cycles and power) The GPU efficiently burst writes all blended pixels from the tile
Slide 8: Qualcomm Technologies' Approach to Binning/Tiling
Qualcomm Technologies’ Approach to Binning/Tiling • The only output of Binning Pass is a compressed 1 bit/primitive ‘visibility stream’ • Avoids write bandwidth for transformed positions and overflow/ • management of transformed VBO.
Slide 9: Tiling Gotcha's
Tiling Gotcha’s • Use ‘Frame Buffer Discard’ to indicate that old data will be overwritten – Otherwise we need to copy (‘unresolve’) the old frame buffer back into the Tile Buffer – ‘Frame Buffer Clear’ works as well – but means we have to waste time clearing the Tile Buffer • Mid frame calls of glT
Slide 10: Bandwidth Saving Technique – Deferred Rendering
Bandwidth Saving Technique – Deferred Rending • ‘Hidden Surface Removal’ or ‘Deferred Rendering’ can avoid rendering occluded pixels – even for non-sorted command streams. • Works by deferring color render until Z value is known – similar to a Z pre- render. • Doesn’t work for all cases -- Alpha Ble
Slide 11: Bandwidth Saving Technique – Forward Pixel Kill
Bandwidth Saving Technique – Forward Pixel Kill • Use ‘Early Z’ test to ‘reach into’ processing pipeline and kill pixels that are still being processed. • Many of the benefits of ‘Deferred Rendering’ • Kills even with back-to-front Ordering • Can kill pixels even if the tile has Alpha Blending. • Ef
Slide 12: Bandwidth Saving Technique – Tessellation
Bandwidth Saving Technique – Tessellation • Supported in AEP/OpenGL ES 3.1 • Provides the ability to render very smooth and high detailed scenes Tessellation OFF Tessellation ON
Slide 13: Bandwidth Saving Technique - Tessellation
Bandwidth Saving Technique - Tessellation 3 Different Possibilities and Bounding Box Extension 1. Tessellate on Binning Pass I. Works – but requires extra bandwidth to Save/Read tessellated Vertices 2. Tessellate on Render Pass I. Saves bandwidth - but requires tessellated primitive to appear in all
Slide 14: Bandwidth Saving Technique - Texture Compression
Bandwidth Saving Technique - Texture Compression • Texture compression is a critical tool for reducing bandwidth - which improves performance and lowers power consumption • All textures containing color data should be compressed using one of the recommended formats below. Original ASTC Compression 2
Slide 15: Render Target Compression
Bandwidth Saving Technique – Render Target Compression • Lossless – means that can be invoked anytime the CPU doesn’t need access. Doesn’t save space -- only memory bandwidth. • End-to-End (understood by both Texture Unit and Display) means displayable Targets can be compressed – and saves bandwidth
Slide 16: Summary
Summary • Bandwidth is becoming the dominant performance and power consumption consideration for mobile graphics • Mobile chips use ‘tricks’ to try minimize bandwidth as much as possible. • Please cooperate with us by using mobile ‘friendly’ algorithms and using compressed surfaces.
Qualcomm Technologies — Technical Presentation (2015) on 6ix