capi snap framework - openpower foundation€¦ · 14-10-2016 · capi snap framework •concept...
TRANSCRIPT
© 2016 OpenPOWER Foundation
CAPI SNAP Framework
Bruce Wile
CAPI SNAP Framework
• Concept initiated at OpenPOWER Summit San Jose in April, 2016
• Two significant directives:1. The framework must make it easy for programmers to call accelerators and
write their own acceleration IP.
2. The framework must be open source to enable continued enhancements and cross company collaboration.
• Framework will be release on github at the end of the year
• Seven OpenPOWER companies actively contributing:
© 2015 OpenPOWER Foundation 2
SNAP = Storage, Networking, and Analytics Programming
CAPI SNAP Overview
© 2016 OpenPOWER Foundation 3
POWER8 Cores
YourApplication
CAPI Framework
Library
YourApplicationYour
Application
APISource,
Action,
Destination
Virtual Address Space
Virtual Address Space
Virtual Address Space
CAPI over
PCI-E
FPGA
FrameworkAction Code
8G Local
DRAM
Network,
Flash,
SAN, …
FPGA
FrameworkAction Code
2TB Flash
Make CAPI easy for developers: API
© 2016 OpenPOWER Foundation 4
With CAPI Developer Kit
(for computer engineers)
With CAPI SNAP Framework
(for programmers)
One function call:
Source
Action
Destination
capi_addr
address
size type flags
capi_addr
address
size type flags
Job:
action/kernel
Return_code
Make CAPI easy for developers: Accelerator Code
© 2016 OpenPOWER Foundation 5
/*Command Interface*/
AXH_CVALID
AXH_CTAG(0:7)
AXH_COM(0:12)
AXH_CEA(0:63)
AXH_CSIZE(0:11)
HAX_CROOM(0:7)
b”1”
x”20”
x”0D70”x”0000000000FB0810”
x”010”
x”40”
/* Command Interface*/
20 21 221F
0D70
010
0A60
0000000000FB0810 0000000000FB0A80 0000000000FB0F6000000000B
080
40
With CAPI Developer Kit
(for computer engineers)
With CAPI SNAP Framework
(for programmers)
Key Dates and Actions
• November – December: Early Partner initial testing
• January, 2017: Framework available via release to github• Initial card for 2TB flash and off-load engines
• For more information and to apply for Early Partner testing:• Email: [email protected]
© 2016 OpenPOWER Foundation 6
© 2016 OpenPOWER Foundation
The 2016 OpenPOWER Developer Challenge
The OpenPOWER Foundation challenged Developers
to port, optimize, scale or create new HPC & Big Data
& Analytics applications on POWER + HW accelerators.
Three competitions include
+ Open Road Test
+ Accelerated Spark Rally
+ Cognitive Cup
© 2016 OpenPOWER Foundation
Antonio Carlos Furtado
University of Alberta
Predicting the frequency
of emergency events to
reduce costs and save
lives
Accelerated Spark Rally
Emergency prediction
on Spark
Grand Prize Winners
Andrei Yurkevich
(Team Lead)
Altoros Labs
Scaling Deep Learning
for automated cancer
detection
Cognitive Cup (Tie for 1st)
Distributed TensorFlow
for cancer detection
Zaid Al-ars
(Team Lead)
Delft University of
Technology
Scaling PairHMM up
and out to advance
NGS
Open Road Test
Scaling up & out a
bioinformatics algorithm
Praveen Sridhar
(Team Lead)
Deep Learning to classify,
order, and sort paintings by
genre
Cognitive Cup (Tie for 1st)
artNet genre classifier