GS-4152, AMD’s Radeon R9-290X, One Big dGPU, by Michael Mantor

AMD RADEON™ R9-‐290X, ONE BIG DGPU MIKE MANTOR – [email protected]

GRAPHICS, COMPUTE & VISUALIZATION ARCHITECT

AMD RADEON™ R9-‐290X, ONE BIG DGPU

Audio Technologies

Specifications

GCN Architecture

Display Technologies

A CONTINUED ERA OF GCN GRAPHICS & COMPUTE

3 | RADEON™ R9-‐290X, ONE BIG DGPU | NOV 2013 | APU13

GRAPHICS CORE NEXT ARCHITECTURE

!  44 Compute Units !  4 Geometry Processors

‒  4 billion primiIves/sec

!  64 Pixel Output/Clock ‒  64 Gpixels/sec fill rate

!  176 Texture Units ‒  176 Gtextured pixels/sec

!  704 32b Load/Store Units ‒  Up-‐to 2.8 TB/sec L1 bandwidth

!  1MB R/W L2 Cache ‒  Up-‐to 1 TB/sec L2/L1 bandwidth

!  512-‐bit GDDR5 memory interface ‒  320 GB/sec memory bandwidth

!  New AMD CrossFire™ technology !  AMD TrueAudio technology !  6.2 billion transistors

‒  438 mm2 on 28nm process node

RADEON™ R9 290X

A NEW GPU DESIGN FOR A NEW ERA OF COMPUTING


! RADEON™ R9 290X GCN GPU contains 4 Shader Engines ‒  Load balanced with each other ‒ Screen partitioning for pixel assignment

! A Shader Engine is a high level organizational unit containing: ‒ 1 Geometry Processor (1 Primitive Per Cycle Throughput) ‒ 1 Rasterizer ‒ 1-11 CUs (Compute Units)

‒  Instruction I$ and constant K$ caches shared by up to 4 CU each

‒ 4 RBEs (Render Back Ends) ‒ Up-to 16 – 64b pixels/cycle per Shader Engine ‒ Up-to 8 – 128b pixels/cycle per Shader Engine

SHADER ENGINE RADEON™ R9 290X GRAPHICS CORE NEXT ARCHITECTURE


!  CU = Basic Building Block of GPU Computational Power ‒  44 compute units (2816 SP IEEE 2008-‐compliant shaders) ‒  Device Flat (Generic) 64b/32b Addressing support ‒ Masked Quad Sum of Absolute Difference (MQSAD) with 32b AccumulaIon and SaturaIon ‒  Precision improvement for naIve LOG/EXP ops to 1ULP ‒  Packed 64Bb pixel color exports – effecIvely double reorder buffering in 64b pixel processing w/access power reducIon ‒  New Local Data Share (LDS) Memory OperaIons

COMPUTE UNIT

Branch & Message Unit Scalar Unit Vector Units

(4x SIMD-16)

Vector Registers (4x 64KB)

Texture Filter Units (4)

Local Data Share (64KB)

L1 Cache (16KB)

Scheduler Texture Fetch Load / Store

Units (16)

Scalar Registers (8KB)

RADEON™ R9 290X GRAPHICS CORE NEXT ARCHITECTURE


!  Flat Address Space (“flat”) instrucIons are new for Radeon™ R9-‐290X and allow read/write/atomic access to a generic memory address pointer which can resolve to any of the following physical memories:

!  Global memory

!  Scratch (“private”) !  LDS (“shared”) !  Invalid -‐ MEM_VIOL TrapStatus

!  Device Flat(Generic) 64b/32b Addressing mode support ‒  FLAT instrucIons support both 64 and 32-‐bit addressing. The address size is set via a mode register (“PTR32”) and a local copy of the value is stored per wave.

‒ The state addresses for the aperture check differ in 32 and 64-‐bit mode

COMPUTE UNIT RADEON™ R9 290X DEVICE FLAT ADDRESS

FLAT


! New Local Data Share Memory OperaIons ‒ F32 Min/Max/Compare Swap ‒ Full complement of LDS remote atomic op with shared memory dual source operands ‒ LDS[Dst] = LDS[addr0] op LDS[addr1]; ‒ Fast remote reducIon operaIons for arithmeIc, logical, Min/Max F32

‒ Read/Write Exchange 96b/128b

COMPUTE UNIT RADEON™ R9 290X LDS OPERATIONS


! SAD = Sum of Absolute Differences !  Critical to many video & image processing algorithms

‒  Motion detection ‒  Gesture recognition ‒  Video & image search ‒  Stereo depth extraction ‒  Computer vision

!  SAD (4x1) and QSAD (4 4x1) instructions ‒  QSAD combines SAD with alignment ops for higher performance and reduced power draw ‒  Evaluate up to 256 pixels per CU per clock cycle!

!  Maskable MQSAD instruction ‒  Allows background pixels to be ignored ‒  Accelerated isolation of moving objects

!  New : 32 bit destination accumulator register ‒ SAD/QSAD/MQSAD U32/U16 accumulators with saturation

2 5

4 0

7 5

4 1

5 4

7 1

9 4

3 1

3 5

4 0

2 5

8 1

5 5

7 1

9 7

1 7

9 3

2 9

1 5

6 8

1 4

6 6

3 2

3 0

3 2

7 4

5 0

9 2

9 9

5 8

4 0

4 7

3 0

2 2

8 2

1 0

7 1

2 0

3 9

3 6

SAD = 59

SAD = 45

SAD = 58

SAD = 22

Closest match

SAD = 22

AMD Radeon™ R9 290x can evaluate 11.26 Terapixels/sec *

* Peak theoreIcal performance for 8-‐bit integer pixels

GCN MEDIA PROCESSING MEDIA INSTRUCTIONS RADEON™ R9 290X


Process and rasterize up to 4 primi\ves per clock cycle

! Updated Geometry Processor Units – Off-chip buffer management improved

–  Larger posi\on & parameter caches

! GS + Tessellation is faster than before…

! However… memory can still be the bottleneck!

– Minimize the number of inputs and outputs for best performance…

! Small GS expansions can be done within LDS eliminating cache/memory traffic!

TessellaIon off

GCN FIXED FUNCTION ARCHITECTURE GEOMETRY Geometry Processor Geometry Processor Geometry Processor

Tessellator Vertex Assembler

Geometry Assembler

Geometry Processor


Geometry Assembler Tessellator Vertex

Assembler Geometry Assembler Tessellator Vertex

Assembler Geometry Assembler

TessellaIon off

Image from Baulefield 3, EA DICE

TessellaIon on

RADEON™ R9 290X


Z/Stencil ROPs Color ROPs

Depth Cache Color Cache

!  Depth/Color/Stencil processing conducted in the Render Back-Ends (RBEs) ‒  4KB Depth Cache

‒ Up to 16 coverage samples ‒ 16KB Color Cache

‒ Up to 8 color samples

!  Radeon™ R9-290X fragment rates ‒ 64 – 64b pixels per cycle ‒ 256 Depth Test (Z) / Stencil Ops per

cycle

* There are 16 RBEs on Radeon™ R9 290x

GCN FIXED FUNCTION ARCHITECTURE RENDER BACK ENDS RADEON™ R9 290X


! There are 4 Geometry Processors on Radeon™ Radeon™ R9 290x

‒  Overall Primitive Rate = 4 prims per clock (ideal)

! There are 4 Rasterizers on Radeon™ R9 290x (4 triangles x 16 pixels = 64 pixels per clock)

‒  Each rasterizer can read in a single triangle per cycle, and write out up-to 16 pixels

! Caveat: tiny (e.g. sub-pixel) triangles can dramatically reduce PS efficiency !  This can cause us to become raster-bound, unable to rasterize at peak-rate!

Geometry Processor


Geometry Assembler

Rasterizer

Scan Converter Hierarchical Z

Render Back-Ends

Rasterizer


Render Back-Ends

Geometry Processor


Geometry Assembler

Rasterizer


Render Back-Ends

Rasterizer


Render Back-Ends

Geometry Processor


Geometry Assembler

Geometry Processor


Geometry Assembler

Command Processor

Compute Units

GCN FIXED FUNCTION ARCHITECTURE RASTERIZER RADEON™ R9 290X


1 MB L2 R/W CACHE !  Upto16 - 64 KB L2 cache partition

‒ Relative to AMD Radeon™ HD 7970 ASIC ‒ Additional 33% last level cache capacity ‒ Additional 33% internal fabric bandwidth

‒  Up to 1TB/s L2/L1 bandwidth @ 1GHz operation

L2 Cache

Memory Interface ! 512-‐bit GDDR5 at 5.0 Gbps

‒ 320 GB/sec memory bandwidth

!  High density interface design ‒ 512-‐bit memory interface occupies ~20% less area than the 384-‐bit memory interface uIlized on the AMD Radeon™ HD 7970 ASIC

‒ Over 20% more total bandwidth ‒ 50% increase in bandwidth per mm2

64-bit Dual Channel Memory Controller








L2 & MEMORY INTERFACE RADEON™ R9 290X GRAPHICS CORE NEXT ARCHITECTURE


Tiled Resources & Partially Resident Textures

! Enables application to manage more texture data than can physically fit in a fixed footprint ‒ Known as: Tiled Resources (Direct3D 11.2) and Partially Resident Textures (OpenGL 4.2) ‒ A.k.a. “Virtual texturing“ and “Sparse texturing”

! The principle behind PRT is that not all texture contents are likely to be needed at any given time ‒ Current render view may only require selected portions of a texture to be resident in memory ‒ Or, only selected MIPMap levels…

! PRT textures only have a portion of their data mapped into GPU-accessible memory at a time ‒  Texture data can be streamed in on-demand ‒  Texture sizes up-to 32TB (16k x 16k x 8k x 128-bit)

! All Tier-‐2 features supported in Radeon™R9 – 290X ‒  Shader LOD clamp & mapped status feedback ‒ Reads from non-‐mapped Ile returns 0 ‒ Min/max reducIon filtering

!  OpenGL extension – GL_AMD_sparse_texture


GCN COMPUTE ARCHITECTURE RADEON™ R9 290X

A NEW CONFIGURATION OF GCN FOR GAMMING AND COMPUTING

AMD Radeon™ HD 7970 GHz Edition

AMD Radeon™ R9 290X Increase

Geometry Processing 2.1 billion primitives/sec 4 billion primitives/sec 1.9x

Compute 4.3 TFLOPS 5.6 TFLOPS 1.3x

Texture fill rate 134.4 Gtexels/sec 176 Gtexels/sec 1.3x

Pixel fill rate 33.6 Gpixels/sec 64 Gpixels/sec 1.9x

Peak Bandwidth 264 GB/sec 320 GB/sec 1.2x

Die area 352 mm2 438 mm2 1.24x

Peak GFLOPS/mm2 12.2 12.8 1.05x


GRAPHICS CORE NEXT

!  8 ASYNCHRONOUS COMPUTE ENGINES (ACE) ‒  Operate in parallel with Graphics CP ‒  Independent scheduling and work item dispatch

for efficient multi-tasking ‒  9 Devices with 64+ Command Queues!

‒  Fast work group boundary preemption ‒  Exposed in OpenCL™

!  Dual DMA engines ‒  Can saturate PCIe 3.0 x16 bus bandwidth (16 GB/

sec bidirectional)




GRAPHICS CORE NEXT

!  ACEs are responsible for compute shader scheduling, synchronization & resource alloc

!  Each ACE fetches commands from cache or memory

!  Tasks have a priority level for scheduling ‒  Background à Realtime

!  ACE dispatch tasks to shader engines as resources permit

!  Tasks complete out-of-order, tracked by ACE for correctness

!  Every cycle, an ACE can create a workgroup and dispatch to the shader engines for round robin for wave launch ‒  Policies:priorities/keep out/limits/reservations etc




GRAPHICS CORE NEXT

!  ACE are independent ‒  But, can synchronize and communicate

via Cache/Memory/GDS

!  ACE can form task graphs ‒ Individual tasks can have

dependencies on one another ‒ Can depend on another ACE ‒ Can depend on part of graphics pipe

!  ACE can control task switching ‒ Stop and Start tasks and dispatch

work to shader engines




! Ace -‐ Independent virtual engine, enabling true GPU MulIprocessor operaIon ! Operates out of assigned virtual memory (All commands, instrucIons, constants, input, output memory space)

! Doorbell hardware feature enables architected efficient queue write pointer update

! Producer/Consumer can be built on OpenCl™ and advanced HSA ! Footprint limiIng with load balancing techniques between concurrent compute ACE threads

ACE F32 GDS

Memory

Init

Dis

patc

h

Con

trol

Dat

a

CS CS Unified Shader

Intermediate

Data

Memory Input Data

Per WG Data

Sync

Spawn

Alloc ACE F32

Dis

patc

h

Spawn

RADEON™ R9 290X FUTURE CS CONCURRENT MULTITHREADING


GRAPHICS CORE NEXT

!  Focus in GPU hardware migrating towards general-purpose compute units

!  Radeon™R9 290x GCN-based ASICs have advanced compute ‒ CP can dispatch compute ‒ ACE cannot dispatch graphics

!  Future will provide interactive compute and graphics

! If you aren’t writing Compute Shaders, you’re not getting the absolute most out of modern GPUs


RADEON™ R9 290X GCN COMPUTE ARCHITECTURE


NEW AMD CROSSFIRE™ TECHNOLOGY -‐ XDMA

!  XDMA is a modern approach to negoIaIng the communicaIon of mulIple GPUs for peak gaming performance with the ultra-‐high-‐resoluIon displays of tomorrow

!  Radeon™R9-‐290X includes a Hardware DMA engine in the AMD CrossFire™ composiIng block ‒ Designed for AMD Eyefinity and UltraHD resoluIons via DisplayPort™ ‒ Allows for direct access between GPU display pipelines over PCI Express® ‒ No external connector required ‒ CompaIble on AMD Catalyst™ frame pacing technologies

‒ No performance penalty versus external bridge


AMD EYEFINITY TECHNOLOGY

SEPTEMBER 2009


EYEFINITY 6 EDITION

APRIL 2010

1


5X1 PORTRAIT MODE

OCTOBER 2011


HDMI® (WITH 4K)

FEBRUARY 2012 DISPLAYPORT™ 1.2

DECEMBER 2011

32”

3840

2160

3840

50”

2160


THE GENERATIONAL LEAP GAMERS HAVE BEEN WAITING FOR

4K

1080p

SD

4X THE RESOLUTION OF 1080p


THE FUTURE OF ULTRAHD

!  We believe the AMD Radeon™ R9 290X will support future third type of UltraHD display that feature Single Stream operaIon at 60Hz

!  This product is ready with support for high pixel rates up to 600 MHz, and will validate as soon as the monitors are available 3840

2160


1 x HDMI®

1 x DisplayPort™ with Multi-Stream Transport

2 x Dual-Link DVI

BUILT FOR USER FLEXIBILITY SUPPORT FOR UP TO 6 DISPLAYS 1


2 x Dual-Link DVI 1 x HDMI® x DisplayPort™

with Multi-Stream Transport HAVE 3 MATCHING MONITORS?

USE ANY THREE OUTPUTS


BUILT FOR USER FLEXIBILITY SUPPORT FOR UP TO 6 DISPLAYS

2 x Dual-Link DVI

1 x HDMI®




1 x HDMI®


2 x Dual-Link DVI



1 x HDMI®


2 x Dual-Link DVI


!  TrueAudio on AMD Radeon™ R9 290X

! Dedicated Audio DSP soluIon for game sound effects

!  Guaranteed real-‐Ime performance and service

!  Programmable pipeline provides arIsIc freedom to game audio engineers

!  Intended to transform game audio as programmable shaders transformed graphics

GCN MEDIA PROCESSING AMD TRUEAUDIO


3D AUDIO DONE RIGHT

Absence of 3D positional data from the game leads to “fake” 3D audio

Unprocessed Stereo Sound

ApplicaIon / Game USB Audio Driver

USB

!  USB headset solutions provide processing at the end of the pipeline ‒  Software & Hardware solutions

Traditional Virtual Surround Sound

AMD & GENAUDIO will bring this experience to gamers

Processed Stereo Sound

Audio Middleware

Effects & Algorithms

AMD TRUEAUDIO

ApplicaIon / Game Audio Assets

!  Spatialization algorithm with 3D positional data from game leads to more accurate 3D audio

Virtual Surround with AMD TrueAudio Technology


SPATIALIZATION / 3D AUDIO REVERBS

AUDIO/VOICE STREAMS MASTERING LIMITERS

GCN MEDIA PROCESSING AMD TRUEAUDIO


THE DRIVING FORCES BEHIND THE AMD RADEON™ R9 290X

NEW GAME CONSOLES

ULTRAHD DISPLAYS

MULTI-MONITOR GAMING

AUDIO IMMERSION

NEXT GENERATION GAMES

Works with all Graphics Core Next GPUs

MANTLE

!  New low level programming interface for PCs

!  Designed in collaboration with top game developers

!  Lightweight driver that allows direct access to GPU hardware

!  Compatible with DirectX® HLSL for simplified porting GCN

Mantle Driver

Mantle API

Graphics Applications

!  GS-4112 – Mantle: Empowering 3D Graphics Innovation

!  Keynote – Johan Andersson, Technical Director, EA

!  GS-4145 – Oxide on Mantle Adoption (Wed 5:00-5:45)


PRODUCT SPECIFICATIONS

AMD RADEON™ R9 290X

Radeon™ R9 290X Stream Processors 2,816

Engine Clock Up to 1 GHz

Compute Performance 5.6 TFLOPS

Memory Configuration 4GB GDDR5 / 512-bit

Memory Speed 5.0 Gbps

Power Connectors 1 x 6-pin, 1 x 8-pin

PCI-E Standard PCI-E 3.0

AMD TrueAudio Technology Yes

API Support DirectX® 11.2 OpenGL 4.3

Mantle


THANK YOU

QUESTIONS?


In this talk, details about how the unique configuraIon and features of the new Radeon™ R9-‐290X are used to deliver a new level of experience for enthusiast gamers. The powerful Graphics Core Next architecture upgraded to drive ULTRAHD 4K resoluIon displays or mulIple mulI-‐monitor configuraIons for great visual effects. The Radeon™ R9-‐290X is equipped with a xDMA engine that uIlizes PCIE for all communicaIon between mulIple GPUs in a system for AMD Crossfire™ operaIons. This soluIon is compaIble with our latest frame pacing updates. The Radeon™ R9-‐290X is addiIonally coupled with AMD’s new TRUEADIO Technology to provide a hardware accelerated programmable audio pipeline, enabling the sound engineer and arIst to significantly engage a parIcipant’s sense of sound in game play.

ABSTRACT


DISCLAIMER & ATTRIBUTION

The informaIon presented in this document is for informaIonal purposes only and may contain technical inaccuracies, omissions and typographical errors.

The informaIon contained herein is subject to change and may be rendered inaccurate for many reasons, including but not limited to product and roadmap changes, component and motherboard version changes, new model and/or product releases, product differences between differing manufacturers, so~ware changes, BIOS flashes, firmware upgrades, or the like. AMD assumes no obligaIon to update or otherwise correct or revise this informaIon. However, AMD reserves the right to revise this informaIon and to make changes from Ime to Ime to the content hereof without obligaIon of AMD to noIfy any person of such revisions or changes.

AMD MAKES NO REPRESENTATIONS OR WARRANTIES WITH RESPECT TO THE CONTENTS HEREOF AND ASSUMES NO RESPONSIBILITY FOR ANY INACCURACIES, ERRORS OR OMISSIONS THAT MAY APPEAR IN THIS INFORMATION.

AMD SPECIFICALLY DISCLAIMS ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR ANY PARTICULAR PURPOSE. IN NO EVENT WILL AMD BE LIABLE TO ANY PERSON FOR ANY DIRECT, INDIRECT, SPECIAL OR OTHER CONSEQUENTIAL DAMAGES ARISING FROM THE USE OF ANY INFORMATION CONTAINED HEREIN, EVEN IF AMD IS EXPRESSLY ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.

ATTRIBUTION

© 2013 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, Radeon, CrossFire, Catalyst and combinaIons thereof are trademarks of Advanced Micro Devices, Inc. in the United States and/or other jurisdicIons. PCI Express is a trademark of the PCI-‐SIG CorporaIon and HDMI is a trademarks of HDMI Licensing, LLC. Other names are for informaIonal purposes only and may be trademarks of their respecIve owners.