motivation for odsa.… · motivation “…the nature of exponentials is that you push them out...
TRANSCRIPT
Motivation for ODSA & Fit into OCP
Dharmesh Jani (DJ)OCP Co-chair and Open Ecosystem LeadFacebook
ODSA
OCP continues to evolve and grow
DataCenter
TripletRack
Battery Cabinet
Freedom Servers
Spitfire Server (AMD)
PowerSupply
Windmill (Intel)
Watermark (AMD)
Mezzanine Card V1
Knox
GroupHug
Open Rack V1
Winterfell
Open Rack V2
Mezzanine Card V2
ColdStorage
Micro Server (Panther)
Wedge
Honey Badger
Leopard
Wedge100
Big Sur
SixPack
Yosemite
Lightning
Backpack
TiogaPass
YosemiteV2
Wedge100S
BryceCanyon
BigBasin
BluRay
TwinLake
OCPNIC3.0
2011 2012 2013 2014 2015 2016 2017 2018
BigBasin V2
Open Accelerator Module
2019
FAV3
Minipack
Minilake
DataCenter
DATA CENTERS
SYSTEMS
SUB-SYSTEMS
MODULES
OCP Accelerator
Module
CWDM4-OCP
Open Domain-Specific Architecture (ODSA) OPEN
CLOSED
WIDETARGETED
CHIPLET APPROACH
Open Domain-Specific Architectures
Architecture
Optimized (Cost and Power)
CPU/GPU Combo
CPU Cores Combo
ODSA: Natural Next Step for OCP
DataCenter
TripletRack
Battery Cabinet
Freedom Servers
Spitfire Server (AMD)
PowerSupply
Windmill (Intel)
Watermark (AMD)
Mezzanine Card V1
Knox
GroupHug
Open Rack V1
Winterfell
Open Rack V2
Mezzanine Card V2
ColdStorage
Micro Server (Panther)
Wedge
Honey Badger
Leopard
Wedge100
Big Sur
SixPack
Yosemite
Lightning
Backpack
TiogaPass
YosemiteV2
Wedge100S
BryceCanyon
BigBasin
BluRay
TwinLake
OCPNIC3.0
2011 2012 2013 2014 2015 2016 2017 2018
BigBasin V2
Open Accelerator Module
2019
FAV3
Minipack
Minilake
DataCenter
DATA CENTERS
SYSTEMS
SUB-SYSTEMS
MODULES
2020
SUB-MODULESCOMPONENTS ODSA
CWDM4-OCP
DataCenter
TripletRack
Battery Cabinet
Freedom Servers
Spitfire Server (AMD)
PowerSupply
Windmill (Intel)
Watermark (AMD)
Mezzanine Card V1
Knox
GroupHug
Open Rack V1
Winterfell
Open Rack V2
Mezzanine Card V2
ColdStorage
Micro Server (Panther)
Wedge
Honey Badger
Leopard
Wedge100
Big Sur
SixPack
Yosemite
Lightning
Backpack
TiogaPass
YosemiteV2
Wedge100S
BryceCanyon
BigBasin
BluRay
TwinLake
OCPNIC3.0
2011 2012 2013 2014 2015 2016 2017 2018
BigBasin V2
Open Accelerator Module
2019
FAV3
Minipack
Minilake
DataCenter
DATA CENTERS
SYSTEMS
SUB-SYSTEMS
MODULES
2020
SUB-MODULESCOMPONENTS ODSA
CWDM4-OCP
ODSA: Natural Next Step for OCP
OPEN
CLOSED
WIDENARROW
CHIPLET APPROACH
Open Domain-Specific Architectures
Architecture
Optimized (Power and Cost)
CPU/GPU Combo
CPU Cores Combo
Start at the beginning…
data source: https://goo.gl/bb6wZW
Moore’s Law: A remarkable journeySERVER
Motivation
….remarkable journey is described by Moore’s Law, Intel co-founder Gordon Moore’s 1967 prediction that the number of transistors we can pack into a microchip would double every 18-24 months.
50 years of exponential growth!!
All exponentials must end…SERVER
Motivation
“…the nature of exponentials is that you push them out and eventually disaster happens”
Some end sooner than others…Dennard’s Law that says as transistors shrink, they get faster, use less power and get cheaper…ended about 10 years back!
data source: https://goo.gl/bb6wZW
Gordon Moore also said…
Cambrian Explosion of Workloads
Early Compute Era
Standard Computation HPC Computation
Standard Computation
Standard Computation
Graphics Computation
Modern Computation
Blo
ck C
ha
in
Co
gn
itiv
e S
ecu
rity
NL
P
Ro
bo
tic
PA
s
Au
ton
om
ou
s V
eh
icle
s
Ma
chin
e L
ea
rnin
g
Ne
ura
l Ne
two
rks
De
ep
Le
arn
ing
Pa
tte
rn R
eco
gn
itio
n
RT
Tra
nsl
ati
on
AI and Machine-learning and data-heavy workloads have exploded in past 5 years and will diversify as new
applications are discovered constantly…
Sp
on
ge
s
Mo
llu
sks
Pro
ka
ryo
tes
Eu
ka
ryo
tes
He
mic
ho
rda
ta
Ech
ino
de
rms
Co
mb
Je
llie
s
Early Precambrian Era
Late Precambrian
Prokaryotes
Eukaryotes
Early Cambrian
Pla
cozo
a
Art
hro
po
ds
Wo
rms
Ch
ord
ata
Bio-Diversity Exploded from single cells into multi-cell organisms during the Cambrian explosion; all
major phylla were established in this transition
Metazoans
All images from Creative Commons
Domain-Specific Architectures
Logic Disaggregation
Improve yield and simplify/relax design requirements
IO Disaggregation
Right functionality in right silicon node
PROVEN EXISTING BUSINESS MODELS
[L. Su, IEDM’17]
CHIPLET VS MONOLITHICS
DIE COST 1X DIE COST 0.59X
I/O
CPU
L1
I/O
CPU
L1 I/O
CPU
L1
I/O
CPU
L1
CPU CPU
CPUCPU
I/O I/O
I/O I/O
L1 L1
L1 L1
CPU
CPU
CPU CPU
CPU
CPU
CPU CPU
Open Domain-Specific Architecture (ODSA) OPEN
CLOSED
WIDETARGETED
CHIPLET APPROACH
Open Domain-Specific Architectures
Architecture
Optimized (Cost and Power)
CPU/GPU Combo
CPU Cores Combo
ODSA Goal: Chiplet Marketplace
PCIe SerDesXSR, Kandou,.. BoW Parallel
AIB, HBM,…
PIPE Interface Adapter Common Abstraction
PCIe Link Layer
SATAC
CIX
Inter-ChipletLink Layer
Intra-ChipletLink Layer
Routing
CacheCoherent
Non-Coherent ISF
Pack
age
aggr
egat
ion
Stac
k
Die
dis
aggr
egat
ion
Stac
k
Inter-chiplet PHY and packaging options
CXL
Tile
Lin
k
Organic substrateInterposers/bridges
HOST
STORAGE
NET
WO
RK
X8 PCIe G3 (64Gbps)40G Ethernet optical
40G Ethernet copper
NFP
FPGA
CPUF
F
QSFP 40G
QSFP 40G
MTP
MTP
F
F
QSFP 40G
QSFP 40G
MTP
MTPDRAM
DRAM
F
F
F
F
PCIe x4 Gen3
ChipletMarketplace
ODSA ActivitiesOCP Form Factors drive Power,I/O
Footprint, Performance
Multi-technologyODSA stack
Reference architectures for:Networking, Storage,Inferencing, Training,Video and Image processing
Business, ToolsWorkflow to assembleProduct
Growth of ODSA
7 10
35
52
65
85
0
10
20
30
40
50
60
70
80
90
OCT DEC JAN MAR JUN SEPT
Cumulative Company Count
0
20
40
60
80
100
120
140
160
180
Jan March June Sept
Attendance in ODSA Events
Source: OCP ODSA Survey
ODSA: A New Server Sub-Project
Extending Moore’s Law:• Domain-Specific Architectures: Programmable ASICs to accelerate high-intensity
workloads (e.g. Tensorflow, Network Flow Processor, Antminer…• Chiplets: Build complex ASICs from multiple die, instead of as monolithic devices, to
reduce development time/costs and manufacturing costs.
Open Domain-Specific Architecture: An architecture to build domain-specific products • Today: All multi-chiplet products are based on proprietary interfaces• Tomorrow: Select best-of-breed chiplets from multiple vendors• Incubating a new group, to define a new open interface, build a PoC
Specifications
SERVER
The Tip of the Spear “Opportunity”
0
100
200
300
400
500
600
2018 2019 2020 2021 2022 2023 2024
Wireless Communications Consumer Electronics Wired Communications
Computing & Data Storage Industrial Electronics Automotive Electronics
SoC SAM by Market Segment
© 2019 IHS Markit
SoC
s bu
ilt w
ith C
hipl
ets
$(M
)
Excludes embedded x86 MPUs marketed as SoCs. Includes Configurable SoCs from PLD Vendors.Source: IHS Markit
Independent research from IHS Markit
Four use cases for chipletsSoCsMPUGPUPLDs
Six Verticals (wireless, wireline, consumer, computing, industrial, automotive)
SoCs shown on the right. Immediate opportunity for an open interface – the tip of the spear.
The End Game
0
500
1,000
1,500
2,000
2,500
3,000
3,500
2018 2019 2020 2021 2022 2023 2024Computing & Data Storage Wired CommunicationsWireless Communications Consumer ElectronicsIndustrial Electronics
Overall Chiplet SAM by Market SegmentAg
greg
ate
Proc
esso
r Mar
ket
built
with
Chi
plet
s $(
M)
0
10,000
20,000
30,000
40,000
50,000
60,000
2020 2025 2030 2035Computing & Data Storage Wired CommunicationsWireless Communications Consumer ElectronicsIndustrial Electronics Automotive Electronics
Extended Chiplet SAM by market segment
Aggr
egat
e Pr
oces
sor M
arke
t bui
lt w
ith C
hipl
ets
$(M
)
An open interface has a huge opportunity
Initially dominated by compute uses case, other market segments grow to dominate
Please Help! Join a WorkstreamJoin Interface/Standards:(Mark Kuemerle/Ramin Farjad/Robert Wang/David Kehlet)
DevelopPackaging + Socket, Dev Board
Provide FPGA IP
Provide ODSA chiplets
Provide PHY technology
Join the PoC, Build fast:(Quinn Jacobson/Jawad Nasrullah/Jayaprakash Balachandran)
Join Business, IP and workflow:(Sam Fuller/Dharmesh Jani)
Develop software Define test andassembly workflow
Provide Chiplet IP
Workstream contact information at the ODSA wiki
Define ArchitecturalInterface
Active ProjectsProject Objective Organizations Participating Recent Results Upcoming Milestones Needs
PHY Analysis PHY requirementsPHY analysisCross-PHY abstraction (PIPE)
Alphawave, AnalogX, Aquantia, Avera Semi, Facebook, Intel, Kandou, Netronome, zGlue,
PHY Analysis paper(published at Hot Interconnect)
PIPE abstraction
BoW Interface No technology license fee, easy to port inter-chiplet interface spec
Aquantia, Avera Semi, Netronome
BoW Interface proposal(published at Hot Interconnect)
BoW specification 0.7End September, 2019
Test chips, Chiplet library supporting interface
Prototype product that integrates existing die from multiple companies into one package
Achronix, Cisco, Netronome, NXP, Samtec, Sarcina, zGlue,Macom, Facebook
Decomposable design flow. Committed schedule End user End user participation~30% funding is open
Chiplet design exchange
Open chiplet physical description format. Ayar, NXP, zGlue, Draft spec ZEF Exchange format draft specification
Link and Network Layer
Interface and implementations –requirements and proposals
Achronix, Avera Semi, Intel, Netronome, NXP, Xilinx
Multi-chiplet test Test requirements for an open-chipletinterface
Chiplet monitoring Monitoring infrastructure for chipletoperation
Business workflow Formalize learnings from prototype effort
Wiki: https://www.opencompute.org/wiki/Server/ODSA, meet Fridays at 8 AM Pacific Time. Please join us.
Engineers from: Achronix, AnalogX, ASE, Avera Semi, Ayar, Cisco, Facebook, Ferric, Intel, Kandou, Macom, Marvel, Netronome, NXP, On Semi, Samtec, Sarcina, Synopsys, Xilinx, zGlue
IF YOU WANT TO GO FAST, GO ALONEIF YOU WANT TO GO FAR, GO TOGETHER
AFRICAN PROVERB