Bruce Wile, CAPI Chief Engineer IBM
Transcript of Bruce Wile, CAPI Chief Engineer IBM
Revolutionizing the Datacenter
Join the Conversation #OpenPOWERSummit
A Year of Growth in CAPI Solutions
Bruce Wile, CAPI Chief Engineer
IBM
Join the Conversation #OpenPOWERSummit
Join the Conversation #OpenPOWERSummit
� Feb 2015
• 2 Accelerators:
� IBM Data Engine for NoSQL
� Algo-Logic L3 Order Book
• Key Message to CAPI Team:
“Go make more accelerators”
• 1 Developer Kit Card
(Nallatech)
� Feb 2016
• Dozens of accelerators
• 3 Developer Kit Companies
(Nallatech, Alpha-Data,
Semptian)
� Each with new cards in the
works
One year later…..Feb 2015 vs. Feb 2016
2
Join the Conversation #OpenPOWERSummit
CAPI Acceleration Types
3
Accelerator
on FPGA
Accelerator Library
ApplicationApplication
ApplicationApplication
Accelerator Building Block
Function API
SAN48-16G
1 U
2 U
3 U
4 U
5 U
6 U
7 U
8 U
9 U
10U
11U
12U
13U
14U
15U
16U
17U
18U
19U
20U
21U
22U
23U
24U
25U
26U
27U
28U
29U
30U
31U
32U
33U
34U
35U
36U
37U
38U
39U
40U
41U
42U
~AC ~AC1 G G8264
121
1-2
840
Power S822LC146GB 15K SAS146GB 15K SAS
Power S822LC146GB 15K SAS
146GB 15K SAS
Power S822LC146GB 15K SAS
146GB 15K SAS
Power S822LC146GB 15K SAS
146GB 15K SAS
Power S822LC146GB 15K SAS
146GB 15K SAS
Power S822LC146GB 15K SAS
146GB 15K SAS
Power S822LC146GB 15K SAS
146GB 15K SAS
146GB 15K SAS
146GB 15K SAS
Application + AcceleratorFull Solution
External IO(available)
Accelerator Types:
Transparent: API is pre-
existing/common such that the
application does not need to change to
take advantage of acceleration.
Integration Required: Application
developer must write code to utilize the
accelerator
Full Solution: Application and
accelerator are part of a complete
solution sold to end customers.
Join the Conversation #OpenPOWERSummit
Catalog of CAPI Accelerators – By Type
4/1/2016 4
Transparent
Erasure Code
GZIP Compression
PairHMM Accelerator
LA Library
Bitwise Encryption
Full Solution
Genomics Processing
Graph Analytics Djikstra
Novara-Fuzzy Text Search
Petascale Indexing
Integration Required
IBM Data Engine for No-SQL
(“CAPI-Flash”)
Fast-Fourier Transfer
Monte Carlo Risk Analysis
CV Library
DNN Library
Key Value Store (KVS)
Dynamic Time Warp Pattern
Match
L3 Order Book
Mood Detection
Object Detection
Object Recognition
Object TrackingWith many more coming…..
Join the Conversation #OpenPOWERSummit
CAPI Development Ecosystem
4/1/2016 5
CAPI Developer Kits available from…. CAPI Frameworks, Examples …..
P8 AFU
DMA Mover
CAPI Framework
Memcopy Example
PSL +
DMA
AFUP8 +
PSL
PSL Sim Engine
AFU
AccDNN-Caffe to CAPI
Join the Conversation #OpenPOWERSummit
IO
Card
6
Database Analytics: Four Paradigms for Comparison
SW Threads
Data
DB2
MEMORY
Analytics
SW Only
Data PathKEY:
HW
Algorithm
Processor Chip
AIO
CardData
DB2
MEMORYFPGA
Analytics
Dev Driver
DD Mem
PCI-E FPGA Acceleration+ Accelerated Analytics on FPGA gains performance for many algorithms
- Device driver overhead and programming difficulty
Processor Chip
B
IO
CardData
DB2
FPGA
Analytics
CAPI FPGA Acceleration
MEMORY
POWER8 Chip
+ Accelerated Analytic on FPGA gains performance for many algorithms
+ Shared Memory model for ease of programming and fast access
+ Performance gains through pipelining of analyzed data (vs. large block
release)
+ Frees processor threads for other work
CA
PI
C
Data
DB2
FPGA
Analytics
CAPI FPGA Acceleration with
Database Integration (Data in Motion)
MEMORY
POWER8 Chip
+ All advantages of CAPI FPGA Acceleration
+ Single data flow into processor IO yields much higher performance
D
CA
PI
Join the Conversation #OpenPOWERSummit
Workload Acceleration of Informix
4/1/2016 7
� Description: This is a project to prove the Value of Power 8 for typical IoT Workload
Acceleration, exemplified by Dynamic Time Warping: a subsequence similarity measurement
algorithm accelerated with FPGA
� Client Value: Faster subsequence similarity search over the historical IoT data than CPU based
computing, balanced in-database analysis(without moving data to a separate analytics
platform) even under consistent data ingestion.
Example: I want to know the times motif like “A” appeared in the history.
AAAA
Join the Conversation #OpenPOWERSummit
Emotion Detection for Retail Analytics
CAPI Bridge
Expression Recognition
Gender
Classify
Face
Detect
Age
Classify
Vision Accelerator Interconnect
Vision Accelerator PlatformOpportunity – Gain valuable shopper
insights by inferring shopper demographic
and behavior patterns from many
strategically placed cameras.
Product
Detection
Gender: Male
Age: 20 – 30
Gaze: 5 seconds
Impression: positive