Ultra Low-power           Searching Microprocessor                   RuChip Pte Ltd Singapore                         Anto...
1. Introduction: the problem, the solution, the customers.   2. Our solution: key technology, key challenges, POC   stage ...
Digital Universe growth    Amount of Digital information in    the world:     •   2006 – 161 exabytes (exa - 1018)     •  ...
Search servers market forecast   Every third server in the world is used for information search   Today the search servers...
The problem    Today’s x86 processors are not fast enough to index the    information as it is being generated    Energy f...
So what is the solution?    New specialized processor with          Ultra-low power consumption          High performance ...
Who our customers will be?Generally:                                    Search start-ups funding over the last years      ...
1. Introduction: the problem, the solution, the customers.   2. Our solution: key technology, key challenges, POC   stage ...
Solution: a new architecture  The solution: a General Purpose  Network Processor (GPNP) for  search applications        Th...
Solving highlighted problems                     Problem                                       Solution (RuChip)The “Power...
RuChip Key Added Value                    NoC                                     Front-end                               ...
Design Targets                    Cost (ASP)                                  Cost (ASP)                                  ...
POC Objectives                                Work packages structure                   Objective                         ...
1. Introduction: the problem, the solution, the customers.   2. Our solution: key technology, key challenges, POC   stage ...
RuChip Supply Chain                                                                 Data centers                          ...
Estimated Market Size for the chipWhen       Real Time Indexing          Market Size (Forecast)                       Appl...
Competitors   There are similar technologies out there which can be   compared to us:         Multicores (picoChip, Ambric...
Disruptive innovation      Deep specialization and customization of the chip          Customization of the Transport Syste...
Go-To-Market Strategy   Product        1-st stage: Transport IP (NP): ASIC + boards        2-nd stage: Multicore chip (GPN...
1. Introduction: the problem, the solution, the customers.   2. Our solution: key technology, key challenges, POC   stage ...
Summary  Problem    • Problem: Insufficient performance and poor energy efficiency of large search engines; The situation ...
1. Introduction: the problem, the solution, the customers.   2. Our solution: key technology, key challenges, POC   stage ...
Key Technology: Transport SystemGoogle bottlenecks:                   Bottleneck                                        De...
Key Technology: protocol     New protocol to effectively manage the (Key, Value) frames                                   ...
GP-NP architecture (part)                                                   NP     GP   Header    Data                    ...
Upcoming SlideShare
Loading in …5
×

Presentation RuChip Pte Ltd

1,767 views

Published on

RuChip presentation for Skolkovo

Published in: Technology, Business
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
1,767
On SlideShare
0
From Embeds
0
Number of Embeds
1,044
Actions
Shares
0
Downloads
6
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Presentation RuChip Pte Ltd

  1. 1. Ultra Low-power Searching Microprocessor RuChip Pte Ltd Singapore Anton Gerasimov Evgeny Kovalev Volkov Mikhail Alexander Blinov Konstantine Stowolosow© Ruchip Pte Ltd confidential 1
  2. 2. 1. Introduction: the problem, the solution, the customers. 2. Our solution: key technology, key challenges, POC stage objectives. 3. Business model. Go-to-market strategy 4. Funding and milestones 5. Summary 6. Annex© Ruchip Pte Ltd confidential 2
  3. 3. Digital Universe growth Amount of Digital information in the world: • 2006 – 161 exabytes (exa - 1018) • 2008 – 450 exabytes • 2011 (forecast) – 1,800 exabytes! • The annual growth rate is expected to be more than 60% Every third server in the world is used for information searchMore information = more processing power: In 2009 there was more than 35 million servers all over the world (US$60B)© Ruchip Pte Ltd confidential 3
  4. 4. Search servers market forecast Every third server in the world is used for information search Today the search servers market size is US$20B a year (source: IDC 2008, Gartner 2009) Text + Text + Images Web Search service evolution: Text Images + Video Demand for new servers: ~1.5M units a year, and growing Servers units 50M Video recognition and indexing 40M 30M Images recognition 20M Real-time and indexing text search 10M 2011 2016 2020 Year© Ruchip Pte Ltd confidential 4
  5. 5. The problem Today’s x86 processors are not fast enough to index the information as it is being generated Energy factor: the necessity to cope with increasing demand for CPU cycles results in huge power consumption And what if the cost of a kWh in 5 years will be $1, not $0.11? Switching to low-power servers based on ARM/Atom CPU having recently appeared on the market (Seamicro, Calxeda) is not a solution Google, Microsoft top executives: “We are not going for ARM/Atom based servers because the software portability overhead is too high”.© Ruchip Pte Ltd confidential 5
  6. 6. So what is the solution? New specialized processor with Ultra-low power consumption High performance due to the use of specialized functional blocks to alleviate the bottlenecks/overheads of a search engine: Data serialization, RPC Data compression/decompression, security Instant large index searching Internal data formats optimized for those of a search engine.Analogy:– an HDTV set-top box has a dozen of special processors andnot an Intel x86 CPU. HDTV data formats are similar to what isused in search engines. It will be playing more and moreimportant role as video indexing becomes a mainstream.© Ruchip Pte Ltd confidential 6
  7. 7. Who our customers will be?Generally: Search start-ups funding over the last years Staurtup Search Distributed <Key,Value> stores and Year Engine Total VC Fundingservices based on them Yahoo! Search 2004 A9.com Amazon.com project Hakai $21 million MSN Search Microsoft projectSpecifically: Ask.com GoodSearch Yahoo project 2005 Search engines: Google, Yahoo, Kosmix Like.com $55 million $48 millionnew search start-ups SearchMe $21 million Live Search Microsoft project 2006 ChaCha $58 million Social networks, social graphs Wikiseek SearchMe project Wikia Search 2007 Blackle.com Google project Hosting for Cloud providers Mahalo $58 million Powerset $21 million Future hosting (Amazon-like, e- Viewzi Cuil $33 million 2008commerce) Boogami VADLO 2009 Bing Microsoft project Average $40 million© Ruchip Pte Ltd confidential 7
  8. 8. 1. Introduction: the problem, the solution, the customers. 2. Our solution: key technology, key challenges, POC stage objectives. 3. Business model. Go-to-market strategy 4. Funding and milestones 5. Summary 6. Annex© Ruchip Pte Ltd confidential 8
  9. 9. Solution: a new architecture The solution: a General Purpose Network Processor (GPNP) for search applications The GP block is based on ARM/Atom cores and ensures low power consumption (x10 better efficiency) The NP block is responsible for alleviating Google’s bottlenecks and intercommunicating with other cores on the board The key functions of GPNP: Data processing Packets parsing Packets routing© Ruchip Pte Ltd confidential 9
  10. 10. Solving highlighted problems Problem Solution (RuChip)The “Power Wall” (power consumption) ARM/Atom “mobile cores” (x10 better)Performance bottlenecks (Google) : New transport system:1) Data serialization, compression, 1) GPNP architecturecryptography 2) New protocol to support data structures2) Instant large index searching 3) Custom instructions extensions, HW acceleration.Portability issue: software stack adaptation HW/SW adaptation layer NP NP NP GP GP GP LOAD BALANCER Software NP NP NP GP GP GP Stack New protocol © Ruchip Pte Ltd confidential 10
  11. 11. RuChip Key Added Value NoC Front-end design Back-end cores design ASIP Chip design Transport Goya micro- System processorThe Transport System is the crucial link in the entire value chain© Ruchip Pte Ltd confidential 11
  12. 12. Design Targets Cost (ASP) Cost (ASP) Existing RuChipPower Performance Power PerformanceConsumption Consumption * Performance validation: Goya chip prototype vs. Amazon Elastic Compute Cloud ** Power consumption estimation: CAD (Synopsys, Cadence) *** Workload is a Nutch/Hadoop framework, embedded version © Ruchip Pte Ltd confidential 12
  13. 13. POC Objectives Work packages structure Objective Description Design of the transport system (NP) – Assess the power consumption and ARCH + RTL + System Model the best/worst performance POC Integration between the GP and NP parts Work with ARM, Seamicro, Intel, Marvell Protocol final design and validation Performance validation compared to Amazon Elastic Cloud Compute Device drivers, firmware API for DMA, Decoders, Parsers. Parsers design for Google, Hadoop, etc. Commercial software stack adaptation Software stack for different perspective applications (other than Google) Transport system concept validation: Protocol parsing; Custom instructions; Encoders/decoders; Security; Transport channels© Ruchip Pte Ltd confidential 13
  14. 14. 1. Introduction: the problem, the solution, the customers. 2. Our solution: key technology, key challenges, POC stage objectives. 3. Business model. Go-to-market strategy 4. Funding and milestones 5. Summary 6. Annex© Ruchip Pte Ltd confidential 14
  15. 15. RuChip Supply Chain Data centers state & corporate Goya Search engines Software global, regional, specialized Server makers Goya & microprocessor System integrators Search Engine Goya specialized board New search engines, Start-ups Boards/chipset Licensing makers Social Networks Facebook, MySpace Goya = Google + Yahoo© Ruchip Pte Ltd confidential 15
  16. 16. Estimated Market Size for the chipWhen Real Time Indexing Market Size (Forecast) Application2011 Text $1B Real time Web monitoring2016 Text + Images $10B People search in the global Web based on an image pattern2020 Text + Images + Video $50B People search the Web for video pattern 1-st Goya chip generation: text indexing 2-nd Goya chip generation: text and images indexing Recognition & Indexing for social services: Buying power of Google (approx.): Year Cost of the new servers One server cost Processor cost New servers in Google / Year 2012 $2300,M $1250 $125 2,100,000 2015 $3900,M $1250 $125 4,200,200 * Assumption: Google spends 33% of CAPEX to buy the new servers ** Assumption: Amortization period for the servers in Google is 4 years© Ruchip Pte Ltd confidential 16
  17. 17. Competitors There are similar technologies out there which can be compared to us: Multicores (picoChip, Ambric, Tilera), Cisco Quantum, Octeon processor (Cavium). Conventional x86 CPU makers (Intel, AMD) represent the biggest threat to our technology Intel’s “Platform 2015” (RMS – Recognition, Mining, Synthesis for Tera-scale computing) Nvidia, IBM, Sun (Oracle) Tesla, Cell etc.© Ruchip Pte Ltd confidential 17
  18. 18. Disruptive innovation Deep specialization and customization of the chip Customization of the Transport System (NP) for the tasks and data structures of Google NP NP NP GP GP GPMain algorithms: Large batch processing LOAD BALANCER (MapReduce) Real time indexing NP NP NP (Dremel) GP GP GP Software: NP – optimization of the GP-GP communication NOSQL Frameworks GP – general purpose part (can be either ARM or Intel) Commercial software stacks (Cloud)© Ruchip Pte Ltd confidential 18
  19. 19. Go-To-Market Strategy Product 1-st stage: Transport IP (NP): ASIC + boards 2-nd stage: Multicore chip (GPNP) + boards Marketing channels Integrators (Novell,..) Direct marketing CASE: Target markets B2B: Search engines, Hosting for NOSQL, Cloud providers, Webmail hosting, Social networks B2C: Mini-search engines (real-time) 100+ Software applications© Ruchip Pte Ltd confidential 19
  20. 20. 1. Introduction: the problem, the solution, the customers. 2. Our solution: key technology, key challenges, POC stage objectives. 3. Business model. Go-to-market strategy 4. Funding and milestones 5. Summary 6. Annex© Ruchip Pte Ltd confidential 20
  21. 21. Summary Problem • Problem: Insufficient performance and poor energy efficiency of large search engines; The situation will deteriorate in several years as the workload increases exponentially (video content indexing) Example: Google’s annual power consumption cost is about $1B. Cost of the new servers is more $1B /Y • Potential customers are: Search engines, Hosting for NoSQL, Webmail hosting, Social networks. Examples: Google, Yahoo, Blekko, Facebook, Yandex, Mail.ru, Microsoft, Baidu, Panguso.com Proposed Solution • The solution: ARM/Atom - based servers and • IP Situation: patent is expected to the end of optimization of the distributed communication system the POC grant. All IP will be concentrated in • Key challenges: Optimization of the transport Singapore. processing system (custom instructions, HW accelerators, new protocol, network processing arch.) Business Model • Market: Brand-new servers for search engines • Disruptive innovation: Deep specialization and and cloud computing (hosting), webmail hosting. customization of the chip through the transport system. • Competitors: Brawny cores companies – Intel, • Revenue model: revenue should come from selling AMD, Sun (Oracle) the chips or licensing the technology.© Ruchip Pte Ltd confidential 21
  22. 22. 1. Introduction: the problem, the solution, the customers. 2. Our solution: key technology, key challenges, POC stage objectives. 3. Business model. Go-to-market strategy 4. Funding and milestones 5. Summary 6. Annex© Ruchip Pte Ltd confidential 22
  23. 23. Key Technology: Transport SystemGoogle bottlenecks: Bottleneck DescriptionOptimization of searching an index Instantly searching an index of more than 100 million gigabytesSerialization, Remote Procedure Call, Data Fast communication between the servers.ExchangeData compression, cryptography Large resource consuming tasks.Scalability to 10 million servers X10 number of servers increase in Google in few yearsNew transport system features: Feature DescriptionHardware support for a custom protocol HW parsers implementation to support the data- structures for Google, Hadoop, Yandex, etc.Hardware acceleration Decoders, CryptographyNew instructions SYNC code detection, fast protocol parsing, fast CRCDifferent transport scenarios To support a very large system scalability© Ruchip Pte Ltd confidential 23
  24. 24. Key Technology: protocol New protocol to effectively manage the (Key, Value) frames HBase HBase File File KV_FRAME KV_FRAME BLOCKS Hbase File structure Protocol Applications BLOCKS Device Drivers Parsers (Google, Hadoop) Protocol will be supported by HW/SW parsers Network abstraction layer* Protocol will natively support <Key,Value> frames for hardware parsing** Protocol will natively support the data structures for different search engines (Google, Yandex,..) © Ruchip Pte Ltd confidential 24
  25. 25. GP-NP architecture (part) NP GP Header Data Table1 Filter Header Filter Application Data Processing Table2 Security MapReduce, Driver Dremel, Table3 Index Search, Speech recognition, Webmail Header Table4 Data Data Security Header Firmware Data Data Security System System Memory System Memory Memory© Ruchip Pte Ltd confidential 25

×