Download presentation
Presentation is loading. Please wait.
Published byMaud Little Modified over 9 years ago
1
Jon Peddie Research - Is the discrete GPU a dinosaur? 1
2
Jon Peddie Research
3
Agenda – What the Roadmaps Say The end of the IGP The introduction of the EPG & HPU The end of the discrete GPU? The big future for discrete GPUs Offsetting factors
4
Jon Peddie Research
5
Strong Market Good Growth
6
Jon Peddie Research PC Platforms
7
Jon Peddie Research Graphics Chips
8
Jon Peddie Research Integrated Graphics Processor
9
Jon Peddie Research GPU Segments PCDesktopDiscreteIGPNotebookDiscreteIGP
10
Jon Peddie Research Intel Processor All that other stuff : USB, CD\DVD, HDD, audio, PCI, parallel serial, 1394, keyboard, IR, etc. Integrated Graphics Processor ( IGP) Unified Memory architecture (UMA) The interface: PCI Express System memory High-speed SDRAM FSB VGA, DVI, Display Port South Bridge Graphics Processor Unit GPU) North Bridge Memory controller Basic Intel PC Architecture
11
Jon Peddie Research IGP Shipments
12
Jon Peddie Research
13
The Integrated Processor Graphics chip
14
Jon Peddie Research First Generation The first versions of this new design can be found in Intel’s Clarksdale i5 processor, a four-core CPU with an IGP in a multichip package. Clarksdale i5 First EPG
15
Jon Peddie Research Introducing the EPG The Integrated PROCESSOR Graphics Clarksdale’s graphics are UMA With the memory controller embedded within the CPU there is a tighter coupling and a higher bandwidth capability. Thus Clarksdale has good performance at nominal screen resolutions (i.e., 1280 x 1024.) (i.e., 1280 x 1024.) Ivy Bridge will be much better, Haswell even more
16
Jon Peddie Research Performance of EPG
17
Jon Peddie Research Desktop EPG Replaces IGP
18
Jon Peddie Research Notebook EPG Replaces IGP
19
Jon Peddie Research PC Segments Workstation (Professional) Mainstream (Consumer & Enterprise) Performance (Multimedia Consumer) Enthusiast (Gamer) Value & Nettop (Enterprise & Consumer) Enthusiast/Workstation (Gamer/Professional) Ultrathin (Consumer) Thin & Light (Consumer & Enterprise) Desktop Replacement (Enterprise & Consumer) Netbook (Consumer)
20
Jon Peddie Research PC Segments Graphics Workstation (Professional) Mainstream (Consumer & Enterprise) Performance (Multimedia Consumer) Enthusiast (Gamer) Value & Nettop (Enterprise & Consumer) Enthusiast/Workstation (Gamer/Professional) Ultrathin (Consumer) Thin & Light (Consumer & Enterprise) Desktop Replacement (Enterprise & Consumer) Netbook (Consumer) I G P Two or more GPUs Discrete GPU I G P Discrete GPU CPU with embedded GPU Discrete GPUs CPU with embedded GPU Discrete GPUs
21
Jon Peddie Research First Fully Integrated EPG 2H’10 AMD shipped samples of its Fusion Llano processor to its OEMs. A fully integrated HPU fabricated in a 32nm process It was a 4 core CPU with a 32 or greater core GPU First GPU to be built in SOI. ATI has demonstrated transistor packing getting 1,600 cores in a 334mm 2 die
22
Jon Peddie Research Llano Block Diagram Core 1Core 2Core 3Core 4 512kb L2 Cache 512kb L2 Cache 512kb L2 Cache 512kb L2 Cache 1024kb L2 Cache 6MB L3 Cache System Request Interface Crossbar Switch DDR3 MCHyperTransport I/O
23
Jon Peddie Research Performance of GPUs Speculated EPGs will be a real inflection point
24
Jon Peddie Research
25
An Inflection Point The EPG is an inflection point One that has been speculated, and worried about for years – has LSI integration gotten to the point that the discrete GPU will be relegated to a small and declining portion of the market? – has LSI integration gotten to the point that the discrete GPU will be relegated to a small and declining portion of the market?
26
Jon Peddie Research Self Perpetuating If that happens, where will the R&D budget come from to develop the next generation of GPUs? The design cycle and cost for a modern GPU is incredible compared to just 10 years ago.
27
Jon Peddie Research Development Costs To get the ROI for that kind of investment the GPU suppliers have to sell not only high-end chips but also midrange derivative versions. If, however, the midrange is satisfied by the EPG based processors, the ROI will take longer and may never be realized due to obsolesce of functionality.
28
Jon Peddie Research Time and Cost to Develop
29
Jon Peddie Research Costs over time The “watershed” concept The average life cycle of a GPU roughly follows the introduction of versions of the Direct X API. DirectX Version Windows Version Release Date DirectX 1.0Windows 3.1Sept. 1995 DirectX 2.095 and NT 4.0Jun. 1996 DirectX 3.0NT 4.0 SP3Sept. 1996 DirectX 4.0n/aNever released DirectX 5.098Jul. 1997Riva 1283D Rage Pro DirectX 6.098 SE and MEAug. 1998Riva TNTRage 128 GL DirectX 7.02000Sept. 1999GF 256Radeon DirectX 8.02000Nov. 2000GF 2 Ultra DirectX 8.1XPNov. 2001 Radeon 8500 DirectX 9.0XPDec. 2002GF FX 5800 Radeon 9700 Pro DirectX 9.0aXPMar. 2003 DirectX 9.0bXPAug. 2003 Radeon X800 DirectX 9.0cXPAug. 2004GF 6800 DirectX 9.0cXPAug. 2005GF 7800Radeon x1300-1900 DirectX 10VistaNov. 2006GF8800 DirectX 10.1Vista SP1Nov. 2007 Radeon HD 2900 DirectX 10.1Vista SP1Feb. 2008GTX 200Radeon HD 4800 DirectX 11Windows 7Oct. 2009 Radeon HD 5800 DirectX 11Windows 7Oct. 2010GTX 480 DirectX 11.1Windows 7Mar 2012GTX 680Radeon 7970
30
Jon Peddie Research Other Concerns As the IGP gained unit market share the investors and press extrapolated the line to predict when all PCs would be IGP based.
31
Jon Peddie Research IGP and GPU Shipments
32
Jon Peddie Research Beware of Soothsayers It’s a silly model and most forecasters knew it The “smart” ones reasoned that Moore’s law would indeed compensate for performance and that functional integration was inevitable and therefore the discrete GPU was a Dodo. Some of them made bets on the shares of the GPU suppliers accordingly, and even used the recession to “prove” they were right.
33
Jon Peddie Research Market Share Issues If GPU shipments stay flat, and EPGs grow at the rate of the PC market, GPU unit market share will shrink. And, depending upon GPU ASP and how the ASP for the EPG is calculated, the GPU market value may shrink. That will make investors very nervous.
34
Jon Peddie Research
35
You can’t build a blivet There is a limit as to how many transistors that can be put into a die There are power envelope limits, package, yield, and heat dissipation limits. You can’t take a 250 watt, 3.6 billion transistor, 1536 core GK104 and cram it in with a 77 watt 1.4 billion transistor Intel Ivy Bridge Core i7 processor Limitations of EPG A blivet is ten pounds of stuff in a five pound bag
36
Jon Peddie Research SoftwareSoftware Software doesn’t stand still If application development stopped Dec 2009 EPG processors would satisfy the needs of 95% of the users. Investors and writers have long seen that as the ultimate conclusion, even before EPGs It wrong!
37
Jon Peddie Research Software Market Issues Hardware has leapt ahead of software Software development is just as time and money consuming as chip design. An AAA first-person shooter (FSP) game takes 3-5 years and $5 to $30 million dollars to produce now It’s a risky “hits-based” business Only a few publishers and studios in the market so the output capacity of the industry is limited.
38
Jon Peddie Research Software Lag Software will continue to improve And use the cycles available to it And stay almost a generation behind HW Occasionally some developers will be in, or almost in sync with the hardware developments.
39
Jon Peddie Research Chicken vs. Egg The chicken and egg dilemma No one wants to develop for a small installed HW base Some ISVs hold back a new release until either the installed base builds up or a competitor announces a new release. This is extremely frustrating for the hardware suppliers.
40
Jon Peddie Research Heterogeneous compute It’s here (at last) Support in three OS This is a significant breakthrough in computer science It will have far reaching, long term, and astounding influence on the world. The economy of scale puts a super computer in the hands of everyone for $5,000 or less.
41
Jon Peddie Research Could this be the Future? CPU Hi speed links to PCIe Bridge Graphics memory High-speed DDRAM South Bridge Graphics Processor (GPU) System DRAM Ray Tracing Processor (RTP) NIC Processor (FPG) Audio Processor (DSP) Larrabee Co-processor Or is the third processor like the third rail?
42
Jon Peddie Research More FLOPS Less $ Cost per FLOP, the GPU used as a vector processor most economical computing element. The GPU can be used in three major application areas: scientific, professional/commercial, and consumer. Most applications will scale by the number of GPU cores available.
43
Jon Peddie Research Destroying things – Investment Protection A discrete-based GPU AIB, adds extraordinary impact to a game. It also offers investment protection to the buyer in that it gives added life to his or her purchase.
44
Jon Peddie Research South Bridge System memory High-speed SDRAM High-speed Links Memory Bus Processor GPU Discrete GPU Graphics memory High-speed DDRAM Frame buffer Hybrid PC Architecture
45
Jon Peddie Research Scalable Graphics SLI – 2 to 3 AIBs Crossfire – 2 to 4 AIBs Lucid Logix – 2 to 4 AIBs - mixed Multiple AIBs for: More performance Physics Ray tracing Investment protection
46
Jon Peddie Research The more you can see … The more you can do - Peddie’s 2nd law In Computer Graphics too much is not enough - Peddie’s 1 st law
47
Jon Peddie Research The big future for discrete GPUs GPUs sell 1.35 to PCs That ratio will increase due to: GPU compute Hybrid Scaling The EPG class products will kill midrange GPUs They can’t come close the high-end
48
Jon Peddie Research The Market Shift to EPG Discrete GPUs
49
Jon Peddie Research Thank you Jon@jonpeddie.com Chasing pixels – finding gems
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.