Scoë Beamer, Chris Baëen, Ajay Joshi (MIT), Krste Asanović

advertisement
Designing Mul+‐Socket Systems Using Silicon Photonics
ScoA Beamer, Chris BaAen, Ajay Joshi (MIT), Krste Asanović, Vladimir Stojanović (MIT) Scott Beamer
Abstract
Technology Comparison
A manycore system will require tremendous
Abstract
bandwidth to memory, and current and projected
electrical interconnects for off-chip
communication may be unable to meet this
challenge. In this work, we use monolithically
integrated silicon photonics to construct a
scalable interconnect between many cores and
memory within a multi-socket system. Our
interconnect also makes it possible to consider
using smaller dies for cost reasons. Our design
can be extended to a general template which can
be configured to a particular die size and number
of sockets.
Abstract
Star Fiber Coupler
Electric
On-chip
Front View
Abstract
Photonic
50fJ/mm
Disintegra4on
Oblique View
Back View
Ribbons
from MCs
150fJ*
S1 S2 S3 S4
S1
Energy
Off-chip
5000fJ
150fJ*
S2
S3
On-chip
100ps/mm
200ps + 10ps/mm
Delay
Off-chip
BW
Density
50ps + 5ps/mm
On-chip
Ribbons
from
Clusters
200ps + 5ps/mm
5 Gb/s/um 160 (320) Gb/s/um
Off-chip
S4
0.2 Gb/s/um
13 (26) Gb/s/um
• Electrical based on 22nm projection
• Simplifies off-chip fiber wiring
• Same number of fibers, but now less ribbons
W/o Star Fiber Coupler
Laser
Compute
Chip
W/ Star Fiber Coupler
Laser Laser
Compute
Chip
Compute
Chip
Ribbon
Compute
Chip
Laser
Silicon Photonics Summary
Chip A
Chip B
Abstract
Off-chip
Laser
Ring
Modulator
Photodetector
2
1
1
2
Ring Filter
• Can transmit 10 Gbps per wavelength (λ)
• Can fit 64 λ per waveguide per direction
Silicon‐Photonic Components
Waveguide
Abstract
Etch Hole
Backend
Dielectric
MOSFET
Poly-Si
Waveguide
Laser
Compute
Chip
Ribbon
Compute
Chip
Compute
Chip
Laser
Physical Packaging
Abstract
DRAM
Design Goals
Laser Laser
Compute
Chip
DRAM
DRAM
• Not all scaling components presented here
Abstract
• Template can handle:
• Cores/die: 16, 32, 64, 128, 256
• Cores/system: 64, 128, 256, 512, 1024
Network Power
Op4cal Power
Abstract
12
DRAM
# cores per die
Laser
• Connect cores to memory controllers
Abstract
• High and uniform bandwidth
• Use photonic advantage to overbuild
• Single photonic hop - best use of technology
• Co-design on and off chip networks to best use
seamless links
STI
Top View
DRAM
Compute
Chip
DRAM
9
DRAM
DRAM
DRAM
DRAM
Star
Coupler
DRAM
DRAM
6
3
Laser
Laser
Compute
Chip
Compute
Chip
Silicon
Substrate
0
•Poly-silicon on top of shallow trench isolation
Resonant Ring
Topology Insights
DRAM
• Rings of the correct dimensions can resonate
• Can use heaters to tune rings (battle variation)
Ring Modulator
DRAM
Core
Core
Core
Core
Core
Core
Core
Core
CoreCore CoreCore
CoreCore CoreCore
L1
L1
L1
L1
L1
L1
L1
L1
L1 L1 L1 L1
L1 L1 L1 L1
L2
L2
L2
L2
L2
L2
L2
L2
Shared L2
Shared L2
Memory
Controller
Memory
Controller
Memory
Controller
Memory
Controller
DRAM
Star
Coupler
• Filter used for wave-division multiplexing
• Use two cascaded rings to get additional
frequency selectivity
256
128
• Optical power 25-30%
of electrical power needed
to make it
64
Connec4ons Off‐Chip
Abstract
Off‐Chip Only
Core Core Core Core
8
MC
Cluster
Cluster
Core Core Core Core
L1 L1 L1 L1
L1 L1 L1 L1
L1 L1 L1 L1
L1 L1 L1 L1
Shared L2
Shared L2
Shared L2
Shared L2
Memory
Controller
Memory
Controller
Memory
Controller
Memory
Controller
Die
1
Die
2
Die
1
Die
2
• When changing number of sockets, all changes
localized to off-chip where easier to change
• Enables scalable system to have uniform
bandwidth
• Group fibers by source (memory controller or
cores)
Power
Fiber
Cluster
Cluster
Conclusions
2x2 Mixers
(8 wide)
8
Cluster
Send/Receive
39%
Power Breakdown
8 waveguides
with 64λ each
MC
Ring Tuning
2%
Laser Generation
58%
Optical
Ribbon
On‐Chip Network Layout
Core Core Core Core
512
DRAM Laser
Optical
Ribbons
Cluster
Ring Filter
Heat Sink
Compute Chip
• Concentration improves link utilization and
reduces serialization latency
Core Core Core Core
DRAM
Concentrated
On‐Chip & Off‐Chip
• Charge injection to change resonant wavelength
DRAM
Par4al Side View
Not Concentrated
resonant light
w/ drop path
1024
Max Cores per System
Concenta4on
Abstract
resonant light
256
128
64
32
16
Laser
Compute
Chip
Air Gap
light not resonant
Template Summary
Laser Power (mW/core)
Silicon Photonics Introduc4on
• Photonic links excel over a distance since
Abstract
conversion cost is constant
• Photonic links are energy efficient off-chip
• Once conversion cost is paid, link can go
seamlessly on and off chip
• Path layout on-chip can greatly impact power
Star
Fiber
Coupler
Since all connectivity is done off-chip, why not
Abstract
build smaller dies?
•Smaller dies will increase yield which decrease
total cost for the same amount of silicon
• A single reusable design will increase volume
• Smaller dies allow for binning on finer
granularity to combat process variation
• A more spread out system will lower the power
density, saving on cooling
16
16
MC
Ribbon
• Multi-socket systems are a place where silicon
Abstract
photonics could offer a large performance
improvement
• Seamless photonic links encourage co-design
of on-chip and off-chip networks
• A photonic interconnect could even allow for
smaller dies (cost incentives)
Cluster
Ribbon
8
Cluster
MC
Cluster
8
MC
• Nested waveguides to avoid crossings
• Each cluster evenly loads nearby waveguides
Future Work
• Multi-hop networks to enable larger designs
Abstract
• Coherence protocols (needed for real systems)
• Going closer to DRAM with photonics
Download