05 ttg labs
TRANSCRIPT
TTG ApptimizerSoftware Autotuning
ToolkitPavel Ivanov, PhD
Co-founder and Deputy CEO, Business Development
How to get the real power of GPGPU?
10- to 20-fold increase in software
performance
10 to 30 percent increase in software
performance
THEORY
PRACTICE
Solution: OPTIMIZE your software !
Let your software to optimize itself !
Manually Time and
resources An endless game
Autotuning Efficiently Once and forever
Traditional approach
Our approach
Multifold performance gain. AUTOMATICALLY !
TTG Apptimizer: as simple as 1-2-3
RESULTS● Highly optimized SW – ↑ 6% in performance● Porting – ↑ 31% in performance FOR FREE !● In general – 1,5- to 3-fold performance increase
GREAT SAVINS FOR THE CUSTOMER !
● Add several lines to your source code● Compile with TTG Apptimizer● Run
Global markets we address● Computational software and services
5
● Big Data
● Game engines and MW
$11B in 2015 / 9%
$7B (6%) est.*
$24B in 2016 / 32%
Team● Maxim Krivov – co-founder, CEO and Head of
development. Certified NVIDIA CUDA Developer. 8 years of experience in HPC ad GPU software development
● Pavel Ivanov – co-founder and Deputy CEO (BD). 21 years in IT industry and project management
● Sergey Grizan – co-founder and leading developer. MS Certifed Professional. 7 years of experience in SW development
● Mikhail Pritula – key developer in customer projects. 8 years of experience in HPC and parallel programming. TopCoder 1611.
We need $1,7M of investments
● To enter into partnerships with vendors and resellers
● To aggressively expand our business onto global market
● To develop new products within TTG Apptimizer family
__global__ void myCudaKernel(){/*Some computations*/}
EnumParameter<int> block(«{128,…, 1024}»);…while (eps > 10e-9){ os.StartIteration(); myCudaKernel<<<N/block, block>>>(); os.FinishIteration();}
Backup slides
Competitive landscapeCompetitors (indirect)
● AMD, Intel, NVIDIA – non-universal solutions
● Caps, GPUSystems, Xcelerit, StarPU,… – no optimization!
Partnership vs Competition
Our advantages
● Software optimizes itself● Universal solution● Shorter time-to-market● Less developer’s skills
Up to 300-fold performance increase*
10Traditional approach (per year)
TTG Apptimizer (once)
0 2 4 6 8 10 12 14 16
Video Streams. Optimization efforts, days
Software optimization: our approach
GPGPU: breakthrough in computations
Porting SW to GPU (semi)automatically
SW performance boosts 10- to 20-fold (expectations)
SW performance boosts 10 to 30 percent (reality)
SW is underoptimized
Optimization for GPUs: an endless race
Our once-and-forever solution:AUTOTUNING
Product and business model Licensing
Cash flowRelated services
GPUs Cost, USD1 19905 2530
10 313020 413030 513050 713075 9130
100 11130
TTG Apptimizer● Dec. 2012 – stable
commercial version● May 2013 – version 1.0.3● First sales● 10 contracts on services● FIIP certificate on SW
Investment needsInvestments
● $1.9M for● TTG Apptimizer versions
for clusters, supercomputers, and clouds
● Marketing, network of partners, sales channels
● Share up to 25%● IRR = 87%● ROI in 4 years
Exit
● Sale of a share to● Investment Foundation● Intel, AMD, NVIDIA
13