What Goes Wrong with Language Definitions and How to Improve the Situation
GPU Acceleration for Financial Services
1. Kinetica – GPU Acceleration for Financial Services
James Mesney, EMEA Systems Engineering Director 1
2. AI will bring significant cost savings and revenue growth opportunities to Banks
AI is transforming Financial Services
Cost Savings and Revenue Growth
Opportunities from AI by 2025 – (Goldman
Sachs)
7+
NEW
ML Libraries
Open Source Released in the last year
3. Deep Learning for Mortgage Risk
Mortgage Risk
https://arxiv.org/abs/1607.02470
2005
2008
2010
2012
• Deep Neural Network
• 20 Years Mortgage Data
• 120 Million Loans
• 3.5 Billion Records
• 2 TB Data
• 300 Explanatory Features
4. GPU Accelerated
In-memory
Scalable
Natural Language
Processing based
full-text search
Geospatial /
location-based
analytics built-in
Analytics in milliseconds
Reduced server sprawl
Commodity hardware
Cloud or on-prem
Real time connectors
to ingest data
Integrated with Hadoop,
Spark, NiFi, Kafka, Storm,
TensorFlow, Caffe, Torch…
+ Rich set of APIs
Data Visualisation
built-in
A Shopping List for Quants and Data Scientists
100x – 1000x faster
than legacy databases
5. 5
WHAT
• Accelerated database designed for parallel
processing across GPU-accelerated hardware
• Ingests billions of records per minute
• Scales to terabytes in-memory
• No indexing or optimizations required
• Integrated visualization engine with native
geospatial support
• Integrated with AI/ML libraries
WHY
• 100-1000x faster than legacy in-memory and
NoSQL databases
• Sub-second response times for billion row
queries
• Ask questions at the speed of thought
• Reduce operational complexity, hardware
footprint and cost
• Converge AI and BI - User Defined Functions
• Built-in Visualisation, Dashboards, Maps
Introducing Kinetica
6. Why Kinetica ?
• Performance
• CPU-only, in-memory databases suffer from lacklustre performance and scalability issues
• Systems struggle to ingest and query simultaneously
• Hadoop can’t deliver acceptable response times for streaming data and near-real-time use cases
• Cost
• Traditional EDWs are expensive and restrictive
• Traditional in-memory databases costly to scale
• Complexity
• Traditional systems require ETL, frequent changes to data models, hardware and software optimizations
• Expensive and hard-to-find skills
• Multiple point-solutions enforces sampling, aggregation, latency.
• Too much “data herding” wastes time.
6
7. The GPU
4,500 cores per device versus
8 to 32 cores per typical CPU
High performance computing
trend to using GPUs to solve
massive processing
challenges GPU acceleration brings high
performance compute to
commodity hardware
Parallel processing is ideal for
scanning/filtering/aggregating
massive datasets
GPUs have thousands of small, efficient cores exceptionally suited to parallel processing.
GPUs are well-suited to compute-intensive workloads, HPC, machine intelligence, deep-learning, AI.
8. Kinetica: Core
IN-MEMORY ANALYTICS DATABASE ACCELERATED BY GPUs
KINETICA
Commodity Hardware
w/ GPUs
Disk
A1 B1 C1
A2 B2 C2
A3 B3 C3
A4 B4 C4
GPU Accelerated
Columnar In-memory Database
HTTP Head Node
• GPU-accelerated, distributed architecture
• Data stored across tiers – VRAM, System memory, SSD,
NVMe, Flash SAN
• Columnar design, relational model…tables, rows, columns
• Column level Dictionary Encoding and Compression
• Native GIS & IP address object support
• Interact with Kinetica through native REST API,
Java/C++/Python API, SQL, ODBC, JDBC, or connectors
• Security
• Authentication : AD/LDAP/Kerberos
• Authorization: Kinetica RBAC
• Audit: Audit log for queries by user and security changes
• Encryption: on disk, 3rd party tool for In-Memory
• SSL/TLS support Typical hardware setup: 2 CPU
Sockets and 256GB - 1TB memory
with 2-4 GPUs per node.
9. Risk Management
MILLISECONDS
STREAM
PROCESSING
ON DEMAND SCALE OUT
IN-DATABASE PROCESSING
MONITORING
Global Positions
Regional Positions
ACCOUNTABILITY
VaR Limits
PRE-TRADE
What-if Scenarios
RISK MANAGEMENT
P&L, Commissions
Sensitivities
Liquidity Risk
Counterparty Risk
Global / Regional
Heads
Desk Managers
Traders
Spot Prices,
Transactions,
Market Risk
Data
External
Transactional
Records
UDF Functions
POSITIONS
ENGINE
CALCULATIONS
PRICING MODELS
RISK
APPLICATIONS
More data available.
Data flowing faster. Many new types of users demanding
sophisticated real-time analysis.
Kinetica
Connectors
10. UNMATCHED
PERFORMANCE
SCALABLE
ACROSS
MULTIPLE NODES
UDF DELIVER 1st
CONVERGED AI AND BI
WORKLOADS
INDUSTRY-STANDARD
CONNECTORS TO DATA
SOURCES & APPS
GPU Accelerated
In-Memory
Streaming and NRT
Commodity hardware
On-premises or Cloud
Scales to 100’s of TB
40x less infrastructure
Machine Learning
AI
In-database
Kafka, Storm, NiFi, Spark
ODBC, JDBC
ANSI SQL/92
API’s for Java, JS, C++,
Python, node.js, REST
Summary - Kinetica GPU Accelerated Analytics