ELEC516 Digital VLSI System Design and Design Automation (spring, 2010) Assignment 4 Reference solution

Transcription

1 ELEC516 Digital VLSI System Design and Design Automation (spring, 010) Assignment 4 Reference solution 1) Pulse-plate 1T DRAM cell a) Timing diagrams for nodes and Y when writing 0 and 1 Timing diagram of pulse-plate 1T DRAM cell Vdd WL PL Vdd Y Write 0 Vdd-Vth Write 1 (vdd) Write 0 (0) Initial store 1 in Ccell Vdd-Vth Vdd-Vth Y Vdd-Vth Write 1 Initial store 0 in Ccell b) From figures in (a), to store 0 in this DRAM cell, voltage of node Y is 0;while writing 1, this voltage changes to V dd V th. The voltage difference between 1 and 0 state increases by ( V V ) compared with normal 1T-DRAM architecture. The dd th profit brought by larger voltage difference includes better noise tolerance, easier design for sense amplifier and smaller leakage etc.

2 c) is pre-discharged to 0 in read operation. When reading 0, since voltage on node Y is zero initially, after asserting WL for read operation V = 0. When reading 1, from charge sharing equation, we ve: ( V V V ) C + 0 C = C V + ( V V ) C dd t dd cell dd cell ( Vdd Vth) Ccell V = Ccell + C d) From equation in c), we can have: ΔV 50mV C > ff ) DRAM with divided bit-line structure a) C is contributed by dummy and DRAM cells as well as the input capacitance of sense amplifier (considering the pre-charge transistor also): Cbit = 56 8 ff + 50 ff + 8 ff =.114 pf (note: if ignoring the pre-charge transistor, Cbit = 56 8 ff + 50 ff + 8 ff =.106 pf) b) Timing diagram is as follows: cell

3 Normally, C is adopted to be close to C, since we want to V d bit pre ( V V ) C = C + C dd th bit d bit close to V dd. c) From charge balance equations, for each side of sense amplifier: C V + C V = ( C + C ) ( V +Δ V) pre c store c pre Cc ( Vstore Vpre) Δ V = C + C To read 0, V store _0 = 0 ; to read 1, Vstore _1 = Vdd Vth ; 1 1 In ideal, we want to Δ Vstore _0 =Δ Vstore _1; Vpre = ( Vstore _1 Vstore _0) = ( Vdd V th) The voltage difference seen by sense amplifier thus can be represented as: 1 Cc 1 Cc Vdd V Vpre = ( Vdd Vth ) 60mV C + C C + C C c ff c c 3) Column Decoder design Consider implementing the design using a hybrid way: (Assume pre-decode x bits) x 4 4 x N = N + N + N = ( x+ 1) + + ( 1) ; dec pre pass tree Since the number of serial pass transistors are restricted to 3. thus x can be,3,4; x= Ndec = 46 x= 3 Ndec = 50 x= 4 N = 96 Thus, the most efficient implementation is pre-decoding bits: The reference schematic is as follows: dec c Pre-decoder

4 4) MSE Processing Element design a) two s complement subtraction on 8bit data: sign Xin Yin 8-bit adder Zout 1 b) One possible design can be as follows: Delay: D = d + d + d + d + d = ns inv 8 bit _ adder inv EX OR 1 bit _ adder 7.6

5 According to timing requirement: T T + D+ T = = 8ns This corresponds to 15MHz. clk Q setup c) PM = SC V f d) For pipelining case, since we reduce the critical path, the maximum clock frequency can be increased. On the other hand, if we maintain the same clock frequency which corresponds to a fixed performance requirement. We can reduce the supply voltage to still meet the delay requirement. From c), by reducing supply voltage, power reduction can be obtained. For parallel case, although SC increases due to duplicating some hardware in the design. The throughput also increases. For the fixed performance requirement, we can reduce clock frequency and supply voltage. From c), we can observe P decreases. e) For design in b), we have : P = ( C + 9 C + 8 C + C + 39 C ) V f M 8 bit _ adder inv xor 1 bit _ adder reg 15 6 PM = ( ) = 16.mW M f) The new design is supposed to be as follows: Pipeline stage added here sign Xin Yin 8-bit adder 4bit 0 1 bit adder C 1 8 XoR array A Delay in stage one is: D1 = tinv + t8 bit _ adder + tinv + txor = = 3.6ns Delay in stage two is: D = t1 _ = 4ns bit adder Maximum clock frequency is : Tclk max( D1, D) + tsetup + tck q = 4.4ns If we still want to run in 15MHz, the new voltage is : Vnew 4.4ns = Vnew =.75V 5V 8ns

6 The new power consumption thus can be formulated: P = SC V f = ( SC + 0 C ) V M _ new new new old reg new + = ( ) mw f 5) Design for testability a) Test vector (X,X,0,0,0) b) The stuck-at-1 fault of node y can also be found by same vector 6) Low Power design a) For transition based encoding, (Assume initial state is ) Data vector Transition based code ( transitions) (7 transitions) (4 transitions) (5 transitions) (1 transitions) (6 transitions) ( transitions) (4 transitions) (6 transitions) Total number of transitions are : E = CV SW = pf V = pj b) For active high coding: Total number of transitions are : E = CV SW = 45 1pf 1V =.5 pj c) For redundant coding:

7 Data vector Redundant code Inv bit ( transitions) (4 transitions) (4 transitions) ( transitions) (3 transitions) ( transitions) (4 transitions) (4 transitions) (1 transitions) Total number of transitions : 6 ; 1 1 E = CV SW = 6 1pf 1V = 13pJ ; 4.% energy saving can be obtained (block diagram of encoder and decoder)