Alpha 21264 Supplement CS 740 Oct. 14, 1998Branch/Jump Target PredictionBranch Prediction LogicBlock DiagramPipeline BehaviorInteresting FactsMeasured PerformanceAlpha 21264 Supplement CS 740Oct. 14, 1998Extracted from“Digital 21264 Sets New Standard”Microprocessor Report, Oct. 28, 1996CS 740 F’98– 2 –Branch/Jump Target PredictionIntegrated into instruction cache•64KB, consisting of 4K lines, each 16B (4 instructions)•Each cache line carries 12b “next line” + 1b “next set” predictor–6.5KB total–Set to predicted target if line contains predicted-taken branch–Set to next sequential line otherwisePredicting Procedure Return Target•Maintain stack of return points for 32 most recent calls•100% accuracy as long as no non-standard returnsAlignment issues•Taken branch best if instruction address of form 4i + 3–Will use other 3 instructions in line•Branch target should have address of form 4i –First instruction in lineCS 740 F’98– 3 –Branch Prediction Logic•Purpose: Predict whether or not branch taken•35Kb of prediction information•2% of total die size•Claim 0.7--1.0% of all instructions are mispredicted branches–But only 1/6 instructions are branches in the first place 4.2–6.0% mispredication rateCS 740 F’98– 4 –Block Diagram4 Integer ALUs•Each can perform simple instructions•2 handle address calculationsRegister Files•32 arch / 80 physical Int•32 arch / 72 physical FP•Int registers duplicated–Extra cycle delay from write in one to read in other–Each has 6 read ports, 4 write ports–Attempt to issue consumer to producer sideCS 740 F’98– 5 –Pipeline Behavior•Misprediction adds 0.1 to SpecInt95 CPI–Yielding 0.5CS 740 F’98– 6 –Interesting FactsVery Complex Chip•15.2 M transistors total•6 M for CPU core–Vs. 4.2M for Intel PentiumProInteresting Memory Subsystem•Have found in past that memory bandwidth is major bottleneck–CPU often “starved” waiting for loads & stores–Not reflected in SPEC benchmarks»Tend to fit entirely in cacheCS 740 F’98– 7 –Measured PerformanceMachines•Alpha 21264 @ 575 MHz–(64+64) KB L1, 4MB L2, 512 MB main memory•Intel Pentium II @450 MHz–(16+16) KB L1, 512KB L2, 64 MB main memorySPEC Results•Alpha–June ‘98–SpecInt95: 30.3–SpecFP95: 47.7•Intel–August ‘98–SpecInt95: 17.2–SpecFP95:
View Full Document