Understanding the Impacts of 3D Stacked Layouts on ILP Vivek Venkatesan, Manu Awasthi, Rajeev Balasubramonian School of Computing, University of Utah BACKGROUND 3D TECHNOLOGY Interconnects within a processor pipeline are known to be a major bottleneck for performance and power in future Drawback : Increased Thermal processors. Wire delays are Vertical 3D stacking of dies Density allows reduction of overall wire-lengths and helps alleviate the performance and power overhead of on-chip wiring. The primary disadvantage is that it results in increased power- RegFile densities and on-chip temperatures. Break and Stack IMPACT OF WIRE DELAYS 3D Folding – Proposed by Puttaswamy et. Al. Wire-latencies predicted to be in 10’s of cycles in future fabrication technologies Studying the impact of wire-delays on performance Observation : Wire-Delay Limited Architectures reinforces the need for interconnect optimization techniques stand to benefit more from a 3D integration Popular belief: Multi-threading hides wire-delays, not technology Proposed Solution : 3D Stacking entirely true! Cluster Technique to alleviate key wire-delays -> Floor-planning Cache Bank (Replicated/Word Interleaved) Arch 1 Arch 2 Arch 3 Reduced Wire Delays and Better Thermal Density !! 3D BENEFITS Best case 3D (Arch 3/WI) performs 12% better than best case 2D (Replicated Cache banks). Better thermal profile : Best case (Arch 3/ WI) has just 10 0 C increase from 2D with maximum performance gains FLOOR-PLANNING Base Case Performance Comparisons Average performance Improvement wrt 2D Centralized 25 25 Floor-planning generates arbitrary layouts of micro- IPC Improvement IPC Improvement 20 20 architectural 15 15 blocks in a processor evaluating each with respect to an 10 10 objective function 5 5 Include delay-criticality information in the objective 0 0 function to Replicated WI Arch 1 Arch 2 Arch 3 keep heavily communicating blocks closer 2D Performance Comparison Best Case 3D Performance 3D floor-planning generates 3D layouts, more potential to exploit closeness in the vertical dimension CONCLUSIONS 3D Technology has the potential to improve processor performance, power and cost 3D wafer bonding traps heat resulting in higher peak and average temperatures Tiled-architectures with long inter-cluster wires stand to gain more from 3D stacking Aggressive cooling capabilities may be required to extract the full potential of 3D Other promising applications of 3D technology include “snap- on” analysis engines, fault-correction engines and stacked memories.