1. Scott Beamer Mehmet Akgul

2. Why do we need CAMs?  Current applications: Networking hardware, i.e. routers Cache Tag Lookup (CPUs)  Design bottlenecks High energy consumption Large area  Desired future applications: Reverse tagged directory for multi-core architectures Our Study: Use EDP as a metric to compare major CAM design methods, demonstrate the pros and cons of individual techniques, and combine them in the optimum way to achieve a low power, high speed CAM design.

3. CAM Cell Operation MATCH LINE SEARCH LINE 1 0 Precharge Match Line 10 Evaluate 0 1 Set Search Line MATCH MISMATCH

4. CAM Cell – NAND Version 10 Precharge Match Line 10 Set Search Line Evaluate 1 Discharge Path 0 1 MATCH MISMATCH 0

5. CAM Power Consumption MATCH LINE SRAM SEARCH LINES Complementary SL → Activity factor = 1 Area, Delay and Energy Penalty Charged full swing every cycle Wasteful in power

6. NOR Cell Analysis ML Energy SL Energy N k

7. NAND Cell Analysis M MXM M M Does not discharge – no energy lost Discharge Path Expected number of NAND cells that switch and consume power

8. NAND & NOR Cell Delay Comparison

9. NAND & NOR Cell Energy Comparison

10. NAND-NOR Hybrid Design NOR Matchlines charge only if all preceding NAND match NAND section discharge when all cells match k N-k kE tot &D tot Find k to optimize EDP

11. EDP for NAND-NOR Hybrid

12. All previous methods focused on ML – What about search line power? Precomputation of 1’s Count 0 0 1 1 0 1 DATA # of 1 01100110 1 0 Search pattern 1 0 01 10 MATCH 0 0 1 1 0 1 Impossible Cases ML Activated 1’s CAM CAM 1’s Count Search Pattern Saves half the search lines Less loading on match lines Reduced Power – Delay – Area

13. Optimized Hybrid Design What is the optimum combination of all these methods? NAND Cells 1’s CountNOR Cells Energy α 1/2 k Slow Half # of SL Overhead Fast High power NAND 1’s Cnt NOR Fast Medium Power NAND 1’s Cnt NOR Slow Low Power 1’s Cnt NANDNOR Slow Sub-optimal design (A) (B) (C)

14. Delay Comparison

15. Energy Comparison

16. Conclusion NAND Cell can achieve very low energy consumption at high delay cost NOR Cell can achieve high speed, but consumes high power To minimize EDP, one should use the optimum number of NAND cells in front of the low swing NOR chain. To maintain ML-SL energy balance, use precombination as a method to reduce the number of search lines by half, and further reducing the activity of ML by 50%. The optimum combination of these individual techniques can be achieved by using NANDs and precoding in front, which conditionally activates the following NOR matchlines, resulting in optimized energy-delay tradeoffs.