Energy-Efficient Computing – Applications in Big … Jain...AppliedMicro’s Aligned IP Portfolio...
Transcript of Energy-Efficient Computing – Applications in Big … Jain...AppliedMicro’s Aligned IP Portfolio...
Energy-Efficient Computing – Applications in Big Data and HPC
Sr. Director of Marketing, Server Solutions
© AppliedMicro Proprietary & Confidential
Connectivity ($1.9B)
AppliedMicro’s Aligned IP Portfolio Strengths
X-Gene™
Fabric
Datacenter Interconnect
Optical Interconnect
+ 64b ARM v8 2-32 Cores @ 2.4 GHz Integrated NIC & Comms Adv. Network Offloads SLIMPro™, mSLIM™
Terabit Coherent Fabric Ultra-Low Latency IO Sharing RDMA Capable up to 1k Nodes
10G BaseT PHYs 10G/25G Backplane Serdes Short Reach Cooper/ Optics 40/100G Uplinks
DC-to-DC OTN Connectivity 10/40/100G Framer/ Mapper Long Reach Optical PHYs for Metro/Long-Haul
‹#› © AppliedMicro Proprietary & Confidential
Content Drivers
PetaBytes Today. ZetaBytes Tomorrow.
0.0
5.0
10.0
15.0
20.0
25.0
30.0
35.0
40.0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Zeta
byte
s 3+ B illion Users Transact via Mobile or Internet in 2014*
500 Smartphones = 1 Server
Datacenter = Information Super Highway
Source: IDC, Gartner*
100 Tablets = 1 Server
Device Drivers
Data Explosion Fuels Datacenter Growth
‹#› © AppliedMicro Proprietary & Confidential
Managed Data Sets
Fragmented Data Sets
Datacenters Undergoing Fundamental Shifts
Consumer Data Corporate Data
Structured Database
Unstructured Database
Distributed Systems
Centralized Systems
Commercial Software
Open-Source Software
‹#› © AppliedMicro Proprietary & Confidential
Why Low Power Servers?
Being Taxed on Power Inefficiencies
Source: James Hamilton, Uptime Institute, AppliedMicro
Power 37%
Others 19%
Servers 44%
Resultant of Large Compute Intensive Cores
Large Power Penalty
Cost of Power = Cost of Server HW
Right-sized solution generates significant TCO savings over life of Datacenter
ARM64b Addresses Mainstream Cloud Workloads
Legacy Enterprise Performance at any cost High power: 200+ W External BoM: chipset,
10G NIC, BMC In-efficient TCO
Performance
Power
X-Gene™
Centerton
E5
E3
Avoton
CloudScaleTM Servers Strong CPU performance Integrated 10G & storage IO Large DRAM memory Sub-40W power envelope General purpose Cloud server
Entry level Appliances Sub-scale performance for
cloud workloads External 10G NIC Application specific; not viable
for general purpose
© AppliedMicro Proprietary & Confidential
7 © AppliedMicro Proprietary & Confidential
Server Platform Evolution
Traditional servers
4U Chassis - 8 servers 2GHz 16
cores per server
- Shared power supply & cooling
- 2.4KW power supply
High Density servers
Back
plan
e
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
XGENE
4U Chassis - 180 servers 2GHz 4
cores per server
4U Chassis - 40 servers 2GHz 16
cores per server
- Shared power supply and cooling - Disaggregation of storage and networking blades
- Common management framework
4 – 6x the density of cores compared to traditional Xeon servers
8 © AppliedMicro Proprietary & Confidential
Server Technology Evolution
• 2 socket Xeon 8C @95W TDP
• External chipset PCIe,
SATA connectivity
• External 10G NIC
Traditional servers
Xeon
D I M M
Xeon
IO Chipset
Fan Control UART B
MC
D I M M
D I M M
D I M M
Mgmt NIC
10G NIC
Low Power servers
CPU
D I M M
D I M M
D I M M
D I M M
Eth PHY
10G PHY
10G MAC
SATA Controller
BMC UART
• Power efficient CPU cores 8-16C @ 20-35W
• SoC approach
integrated SATA, PCIe …
• Integrated 10G significantly reduce latency
Optimized compute for datacenter workloads
High Density Compute Clusters
• IO sharing within the rack networking, storage, PCIe devices
• Flexible topology scales
to 1000’s of nodes
• Optimized IPC – RDMA over 10GE (RoCEE)
CPU
DIMM
CPU
DIMM
CPU
DIMM
CPU
DIMM
Scale Out Fabric
9 © AppliedMicro Proprietary & Confidential
Next-Gen Server Concepts
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
10G 10G
10G 10G
X-G
ene
100G uplink APM Gearbox
10G outbound ports
Backplane
Distributed 10G L2/ L3 switch on each server
6x10G + 2x10G switch ports per server
3D Torus fabric Low latency RoCEE
communication between nodes
100G chassis uplink
10 © AppliedMicro Proprietary & Confidential
Next Gen Intra-Rack Concepts
100G ToR switch
Compute Chassis 1
Compute Chassis 2
Compute Chassis N
Compute Chassis 1
Compute Chassis 2
Compute Chassis N
Storage Chassis 1
Storage Chassis 2
Storage Chassis N
100G
100G
100G
100G
100G
100G
100G
100G
4x25G
4x25G
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
10G 10G
10G 10G
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
10G 10G
10G 10G
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
X-G
ene
10G 10G
10G 10G
X-G
ene
3-
port
100G
sw
itch
APM
Gearbox
APM
Gearbox
10 x10G
GroupHug Chassis Switch
Multiple chassis daisy-chained at 100G; one 100G to ToR
3port 100G switch chassis-to-chassis and chassis-to-ToR communication at 100G
Low latency 100G path to Flash/ SSD storage
© AppliedMicro Proprietary & Confidential
X-Gene™:
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
SATA
Storage I/F
PCIe
Comms I/F
10G I/O
Networking I/F
Coherent Interconnect
Network
Accelerators
Offloads
Multi-Channel DDR3
Memory
High Single Thread Performance
Unique Mixed-Signal IP: 10/40/100G Server on a Chip™ Integration
First to Market
Server Class 64bit Architecture
© AppliedMicro Proprietary & Confidential
X-Gene™:
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
L2 Cache
ARM 64-bit
L1 D
L1 I
ARM 64-bit
L1 D
L1 I
SATA
Storage I/F
PCIe
Comms I/F
10G I/O
Networking I/F
Coherent Interconnect
Network
Accelerators
Offloads
Multi-Channel DDR3
Memory
High Single Thread Performance
Unique Mixed-Signal IP: 10/40/100G Server on a Chip™ Integration
First to Market
Server Class 64bit Architecture + RAS
Samples Availability: Server Availability:
Q1’13. 2H’13.