Developing the Bifrost GPU architecture for mainstream graphics

Transcription

1 Developing the Bifrost GPU architecture for mainstream graphics Anand Patel Senior Product Manager, Media Processing Group ARM Tech Symposia India December 7 th 2016

2 Graphics processing drivers Virtual reality extending into mainstream devices Augmented realities, virtual spaces and immersive videos Vulkan showing great momentum Unity releases Vulkan renderer preview Demand for high quality screens and cameras Smartphone resolutions are rapidly increasing 2

3 Previous Current ARM Mali GPUs and their architectures Mali-G71 Bifrost Mali-G51 Bifrost Mali-470 Utgard Mali-T860/880 Midgard Mali-T820/830 Midgard Mali-400/450 Utgard Performance Area-efficiency Ultra-low-power 3

4 Bifrost GPU design Driver Software Job Manager Shader Core 0 Shader Core 1 Shader Core 2 Shader Core N Control Fabric Tiler MMU L2 Cache Segment L2 Cache Segment L2 Cache Segment AXI Memory Bus AXI Memory Bus AXI Memory Bus 4

5 Mali-G51 scalable and area-efficient design Driver Software Up to 3 dual-pixel shader cores supported Job Manager Single-pixel Shader Core Dual-pixel Shader Core 0 Dual-pixel Shader Core N Control Fabric Tiler MMU L2 Cache Segment L2 Cache Segment AXI Memory Bus AXI Memory Bus 5

6 Dual pixel shader core MP2 and higher configurations use a dual-pixel shader core Increases area efficiency Delivers 2 textured pixels per clock cycle (2-tppc) Configuration Shader core configuration Architecture Execution engines Texels / Clk Pixels / Clk MP1 1x Single core MP2 1x Dual core MP3 1x Dual core + 1x Single core MP4 2 x Dual core MP6 3 x Dual core

7 ZS Memory Bifrost core design Compute Frontend Fragment Frontend Quad Creator Quad Creator Execution Engine 0 Execution Engine 1 Execution Engine 2 Quad State Quad State Quad State Quad Control Quad Manager Message Fabric Load/store Unit Attribute Unit Varying Unit Texture Unit Blender & Tile Access Depth & Stencil To L2 Mem Sys To L2 Mem Sys Tile Memory Tile Writeback To L2 Mem Sys 7

8 ZS Memory Mali-G51 development focus Compute Frontend Fragment Frontend Quad Creator Quad Creator Execution Engine 0 Execution Engine 1 Execution Engine 2 Quad State Quad State Quad State Quad Control Quad Manager Message Fabric Load/store Unit Attribute Unit Varying Unit Texture Unit Blender & Tile Access Depth & Stencil To L2 Mem Sys To L2 Mem Sys Tile Memory Tile Writeback To L2 Mem Sys 8

9 Texture unit The texture unit has been re-designed for Mali-G51 Delivers 1 texel/clk for Single core and 2 texels/clk for Dual core Designed for increased tolerance to high memory system latency Pipeline extensively re-designed to reduce effective pipeline length Reduces die area and power consumption YUV texture sampling performance increased Up to 3x the performance 9

10 Focus on mainstream content Instruction set rebalanced for graphics workloads Added new instructions for shader functions used by more recent graphics apps Tailoring capabilities to mainstream configurations Tiler Memory management unit (MMU) 10

11 AFBC 1.2 ARM frame buffer compression (AFBC) has been updated to version 1.2 DDR-optimized data order for headers and body buffers Improving GPU performance in bandwidth limited scenarios Improving DPU performance in rotation use cases Optimized compression for constant color blocks Full backwards compatibility with AFBC 1.1 specification 11

12 Message preload Work is requested from fixed-function units by sending messages across the message fabric Messages are normally generated by specialized instructions in the execution engines The quad manager can now send messages directly to the varying unit instead of via the execution engines Occurs when the early instructions are varying or varying-texture loads Increases the performance of common fragment shaders 12

13 Area and performance Measured on Mali-G51 Beta candidate RTL 13

14 Power Measured on Mali-G51 Beta candidate RTL 14

15 System level performance advantages Cortex-A53 CPU Mali-V61 Video processor Mali-G51 Graphics processor Mali-DP650 Display processor AFBC Trustzone The premium quality, protected, low-latency, low-power multimedia pipeline 15

16 Mali-G51 status RTL released to lead partners Working towards production-quality RTL in December SoCs expected in 2017, devices expected in 2018 Up to 1.6x more efficient* Mali-G51 publically launched at 2016 ARM Tech Symposia The first Bifrost GPU for mainstream devices Up to 1.6x more area and energy efficient The smallest Vulkan enabled Mali GPU For more information read the launch blog 16 *compared to Mali-T830, on same process node under the same conditions

17 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 2016 ARM Limited