Ultra Low Power GPUs for Wearables

Transcription

1 Ultra Low Power GPUs for Wearables Georgios Keramidas January 2015

2 The Company Who we are? Think Silicon is a privately held company founded in What we do? Development of low power GPU IP semiconductor cores for mobile/embedded devices. Market Focus is the broader IoT and specifically the Wearable market. Our Mission Support and collaborate with our customers to create mutual and enduring values in each phase of the project.

3 Engineering gteam 70 years cumulative industrial i experience 20 Chip Designs Multidisciplinary Team IP portfolio Patent Pool 40 Publications Tier1 OEM under NDA

4 MARKET EVOLUTION

5 Past Decade

6 Display Evolution Smart Watches

7 Future Visualize data from various sources (Internet of Things paradigm)

8 Enabling Technology Graphics Processing Units (GPUs)

9 Moore s Law in Mobile GPUs GPU performance GPU Power Display Resolution Smartphone G 10 years ago 5 years ago NOW in 5 years in 10 years GPUs Applications in next 5 years will need 4-5x the current GPU perf. Display resolution exponential increase (4K displays are here) BUT Power roughly under the same power budget (few hundreds milliwatts) IoT GPUs: few mwatts (< 3 mw) are devoted to graphics IoT GPUs need multi-level power optimizations

10 GPU Challenges Performance Time to market Power consumption

11 Coming years: Wearable Market Think Silicon focuses is on the broader Wearable market with estimated 700 Million shipped devices in the next five years. Worldwide spending on wearable technology will reach $1.4 billion 2014 and by 2018 is predicted to grow to $19 billion.

12 GPU Challenges Power Consumption Time to market Performance

13 GPU Deadlocks Porting GPUs from the high-end phones to the Wearables failed it was reported that at least 30% of the Samsung Gear watches sold by the U.S.-based chain Best Buy were being returned by unsatisfied customers

14 Application Processor NOW Multicore CPU Multicore GPU fully-programmable Video Encoder Video Decoder Image Processors Camera Processors ASIC/ASIP (semi or not) programmable Vision Processors Market needs SMARTER mobile GPUs

15 Application Processor FUTURE Multicore CPU Multicore GPU fully-programmable Video Encoder Video Decoder Image Processors Camera Processors Vision Processors Multicore Multipurpose Multidomain GPUs Mobile GPUs Media Processors & GPGPUs

16 Mobile GPU Challenges Performance Power Consumption Low Die Area Latency Wall Memory Wall Frequency Wall Adaptability to new Standards Time to Market Performance per mm 2 /Watt/$

17 TECHNOLOGY / PROPOSITION

18 product range Portfolio Ultra lowpower Graphics IP Think Series ThinkLCD ML Display Controller Nema Series Think2.5D DirectFB Graphics Engine Nema t OpenGL ES1.1 Nema CL OpenCL 2.0 ThinkVG OpenVG 1.1 Nema GL OpenGL ES 3.0 Graphics Processing Unit High Rich 3D Graphics Vision Processor GPGPU Mid Vector Graphics 2D Graphics Low Text + BMP perf forman nce

19 Opportunity Nema GPU series Built for the Wearables Power budget (single-digit milliwatts operations) Small Display Sizes (e.g., Smartwatches t displays) Software (various light OpenGL operations) Available Memory (libs and images) Memory Bandwidth (various types of compression)

20 Nema series NEMA has been specifically developed for the WEARABLE market! Nema VP OpenVX Nema GL OpenGL ES 3.0 1Q14 Nema CL OpenCL 2.0 4Q14 Nema t OpenGL ES1.1 Low End GPGPU GPU Product Category High End GPU Vision/Imaging/vi deo Processor

21 First Step Nema t,, a tiny GPU OS: API: Interface: RTOS, uclinux, DirectFB, ugfx, OpenGL ES1.1 AMBA AXI/AHB Multicore Architecture: VLIW Core (unema) Multithread, Proprietary ISA Ultra low power Rasterizer Texture Unit: Cache with software Prefetching, 4:1 Texture Compression Real-time 6:1 FrameBuffer Compression 3D support (z-buffer operations) in software Application area: Wrist devices (Smart watch Medical Fitness non Application area: Wrist devices (Smart-watch, Medical, Fitness, non Wearable/low power etc.)

22 First Step Nema t competitive advantage Extends device (e.g., Smart-watch) battery lifetime up to 5 times! Compared to the available IoT GPUs Nema n0 Delivers more performance/mwatt Consumes 4x to 10x less power Occupies 4x less silicon area Saves 3x power consumption from DDR memory accesses (US patent pending - Proprietary Frame buffer compression ) Reduces SoC to Display data traffic (US patent pending)

23 Device Power Consumption Real life scenario Smart Watch: Display 30fps Processor DDR Memory 512MB Battery: 320mA Battery Typical use case Device power consumption * = + SoC power consumption + Memory power consumption + SoC to Display traffic *affected from the GPU

24 System Level Power Consumption LP-DDR Sw library Frame Buffer GPU FB read/ write traffic DDR GPU texture read traffic Mem. Controller CPU SoC Graphics Rendering GPU LCD controller Display interface traffic Display

25 Nema t power savings LP-DDR Sw library Frame Buffer GPU FB read/ write traffic DDR GPU texture read traffic Eliminated Eliminated Power off External DDR! Mem. Controller CPU SoC Graphics Rendering Nema t0 ThinkLCD ML Frame Buffer Sw library SRAM Display interface traffic Reduced Display On-chip TS Compressed Framebuffer Tiny software library (compressed images) Reduced using Selective FrameBuffer Update

26 Device Power Consumption Savings: Nema t Real life scenario Smartwatch battery life extension: 5 times! Memory Power off external Memory Active Power 98.1% Less Active Power 60.9 SoC to Display GPU 0,6 0,7 3,0 3,8 Memory: 100% (less) SoC D: 84.2% (less) GPU: 76.7% (less) (Less is better) NEMA Competition mwatt 0, Memory Power off external Memory Leakage Power 97.4% Less Leakage Power 42,0 GPU 0,07 0,3 Memory: 100% (less) GPU: 76.7% (less) (Less is better) 0,01 0, mwatt NEMA Competition

27 Nema GL Product Nema GL What? Multi core, ultra threaded GPU OS: Linux, Android API: OpenGL ES 3.0 Toolchain: CLang, LLVM, GLSL-LLVM IR (in-house) Why? Low power consumption Extremely small silicon footprint Proprietary Compression Application area: Smart Glasses Head Up Displays Smart Phones 3rdGen Smart Glasses, Head Up Displays, Smart Phones 3rdGen, Tablets 3Gen, Navigation/Tracking-system.

28 NemaGL Special Features Product Nema GL + Nema CL Feature: Compression in Framebuffer Texture data Z-buffer compression Feature: DVFS in Memory-driven di core DVFS Different number of threads/core Feature: Faster single thread performance No Divergent Thread performance (inherit it by design- MIMD not SIMD) (valuable in GPGPU tasks) Compiler driven interlocks & prefetching

29 Company Info HQ Sales & Marketing url: Patras Science Park Ulli Mueller Rion Achaias Toronto Greece Canada info : info@think-silicon.com ili Tel: Fax: Tel: u.mueller@think-silicon.com