Network-on-chip is critical for performance. Network-on-chip is critical for performance.
Electrical signaling - Repeater insertion needed
- Bandwidth density (up to 8 Gbps/um) [Chang HPCA‘08]
Nanophotonics - Bandwidth density ~100 Gbps/ μm !!! [Batten HOTI’08]
- Generally distance independent power consumption
- Speed of light low latency
- Propagation
- Switching [Cianchetti ISCA’09]
Selective Selective - Couple optical energy of a specific wavelength
Modulation & detection Modulation & detection - ~100 Gbps/μm bandwidth density [Batten HOTI’08]
Power Cost Power Cost - Ring heating
- Laser Power
- E/O & O/E conversions
- Distance insensitive
For short links (2.5mm)
Design an architecture that differentiates traffic. Design an architecture that differentiates traffic. - Use electrical signaling for short links.
- Use nanophotonics only for long range traffic.
What do we gain? - Low latency
- High bandwidth density
- High power efficiency
- Localized arbitration
- Scalability
Motivation Motivation Architecture of Firefly Evaluation Conclusion
Concentration Concentration - 4 cores share a router
- 16 routers
Concentration Concentration Clusters - Electrically connected
- Mesh topology
- 4 routers per cluster
- 4 clusters
Concentration Concentration Clusters Assemblies - Routers from different clusters
- Optically connected
- Logical crossbars
Clusters
Single-Write-Multiple-Read (SWMR) [Kirman’06] (CMXbar†) Single-Write-Multiple-Read (SWMR) [Kirman’06] (CMXbar†) - Dedicated sending channel
- Multicast in nature
- Receiver compare & discard
- High fan-out laser power
Multiple-Write-Single-Read (MWSR)[Vantrease’08] (DMXbar†) Multiple-Write-Single-Read (MWSR)[Vantrease’08] (DMXbar†) - Dedicated receiving channel
- Demux to channel
- Global arbitration needed!
Goal Goal Proposed design - Reservation channels
- Multicast to reserve
- Destination ID
- Packet length
- Uni-cast data packet
Virtual-channel router Virtual-channel router - Added optical link ports and extra buffer.
Routing Routing - Intra-cluster routing
- Traversing optical link
Clusters Clusters - Short electrical links
- Concentrated mesh
Assemblies Benefits
Motivation Motivation Architecture of Firefly Evaluation Conclusion
Cycle-accurate simulator (Booksim) Cycle-accurate simulator (Booksim) Firefly vs. CMESH, Dragonfly† and OP_XBAR
Throughput Throughput - Up to 4.8x over OP_XBAR
- At least +70% over Dragonfly
Reduced hardware by partitioning Reduced hardware by partitioning Throughput impact Locality - 34% energy reduction over OP_XBAR with locality
α is heating ratio and β is laser ratio. α is heating ratio and β is laser ratio. Firefly favors traffic locality.
Technology impacts architecture - New opportunities in nanophotonics
- Low latency, high bandwidth density
- Tailored architectures needed
Firefly benefits from nanophotonics by providing - Power Efficiency
- Hybrid signaling
- Partitioned R-SWMR crossbars Reduced hardware/power
- Scalability
- Scalable inter-cluster bandwidth
- Low-radix routers/crossbars
Dostları ilə paylaş: |