14
©2016 MapD. 1 The Promise of GPU Analytics (or why GPU is the new CPU) April 7, 2016 Todd Mostak, CEO, MapD Technologies

The Promise of GPU Analytics (or why GPU is the new CPU)on-demand.gputechconf.com/...mostak-the-promise-of-gpu-analytics-o… · The Promise of GPU Analytics (or why GPU is the new

Embed Size (px)

Citation preview

©2016 MapD. 1

The Promise of GPU Analytics (or why GPU is the new CPU)April 7, 2016Todd Mostak, CEO, MapD Technologies

©2016 MapD. 2

Real-time decisions require highly interactive querying & visualization of big datasets.

But…

Pain point: Impossible on today’s CPU in-memory DBs due to low compute/memory bandwidth.

©2016 MapD. 3

Why Databases Moved In-Memory

• 2005 – Databases move off disk– Vertica, SAP HANA, ParAccel,

Vectorwise– Driven by cheap memory– Dovetails move to column stores

for analytics• Applications see 30-100X

speedups– Mirrors bandwidth difference

between RAM and disk

©2016 MapD. 4

What’s next: GPU In-Memory Databases/Analytics

• GPU memory density increasing, price decreasing

• GPU memory bandwidth rapidly increasing

• High GPU compute bandwidth

• Promises another 30-100X

©2016 MapD. 5

The Chasm

Broadwell Xeon (2-sockets)

Pascal P100 (8 –cards)

Compute (TeraFLOPS, SP)

~2-4 84.8

Memory Bandwidth (GB/sec)

150 5,760

©2016 MapD. 6

GPU processing

The Chasm

20 cores

CPU processing

39,936 cores

©2016 MapD. 7

MapD: Leveraging Multiple GPUs per Server

Column-StoreDatabase &Rendering Engine

Up to 16 Nvidia GPUs per server

• Fast: 100X+ quicker queries• Leverages GPU rendering for live visualizations of

billions of data points• Less hardware, energy, space, maintenance

©2016 MapD. 8

MapD vs Leading Databases

©2016 MapD. 9

MapD Demo

©2016 MapD. 10

The Need for Speed

How MapD achieves its speedups

• Leverages compute and memory bandwidth of multiple GPUs per server

• Partitions and caches hot data in GPU RAM

• Runs on both CPU and GPU

• Dynamically compiles queries into CPU/GPU Code

• Vectorizes query execution

©2016 MapD. 11

MapD frontend• Complex viz, geoviz, charts

Where MapD sits

Principal data store• MemSQL, HDFS or other DB

GPU database• Up to 192 GB memory

GPU-rendered

Tableau or 3rd party viz

ODBC/Thriftconnectors

Non-graphical output

Database

General DataVisualization

Massive DataVisualization(billions of rows)

©2016 MapD. 12

Why now?

Technology

• Higher GPU density + memory sizes now handle most datasets

• CPUs hitting Moore’s law

• GPUs becoming mainstream in enterprise

Market

• Data volumes/velocity exploding

• Heightened need for real-time decision making

• Surge in sensor and geo data

©2016 MapD. 13

Prepared for In-Q-Tel

October 2015

Thank you

©2016 MapD. 14

The Query Pipeline

GPU

SQL Exec Pipeline

SQL Exec + Render Pipeline

MapD web client

SQL + render

SQL-onlySQL-only or SQL + render

Image tilesor Video

Resultstable

Database

Frontend