CT memories. Flash scaling overview

25/02/2010 3D Flash/CT memories Daniele Ielmini DEI - Politecnico di Milano, Milano, Italy [email protected] Flash scaling overview 1. Flash sc...
Author: Rudolph Clarke
38 downloads 1 Views 2MB Size
25/02/2010

3D Flash/CT memories Daniele Ielmini DEI - Politecnico di Milano, Milano, Italy [email protected]

Flash scaling overview 1. Flash scaling:

T. Kamaigaichi, et al., IEDM Tech. Dig. 2008

www.micron.com Feb. 26, 2010

2. Evolutionary scenario:

H. Tanaka et al., VLSI Symp. 2007

3. Paradigm shift

M. Lee, et al., IEDM Tech. Dig. 2007 D. Ielmini, "Non volatile memories" – 3

2

1

25/02/2010

3D memories • 2D memory stacking: – NAND stacking (Samsung, 2006) – TFT NAND (Macronix, 2006)

• 3D charge trapping memories – Bit cost scalable (BICS) memory (Toshiba, 2007) – Terabit Cell array transistor (TCAT) memory (Samsung, 2009)

• Resistive-based crossbar arrays Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

3

Multichip package • This is a ‘fake’ 3D technology: the cost is not reduced (cost is proportional to manufactured area, not package area), only the package area/thickness is affected (attractive for handsets, MP3 players, etc.) • e.g. the Samsung 64GB moviNAND measures 1.4 mm in height, consists of 16 30nm-class 32Gb MLC NAND chips and a controller. The 17-die stack was achieved by using 30-micron thick chips (0.03*17 = 0.51mm) and advanced package technology Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

4

2

25/02/2010

3D NAND • Stacking of NAND layers (not chips) in a single chip • Motivation: further cost reduction

EUV

• In fact, cost scaling slows down because of lithography limitations (immersion litho, double/triple/quadruple patterning, or EUV with huge investments and reduced throughput) • 3D allows less aggressive F scaling  better reliability • Stacking (and 3D) has better costs by avoiding advanced litho tools and adopting relaxed F Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

5

3D NAND

• 2nd layers is SOI with epitaxial Si on ILD • 63 nm TANOS technology, erase by source (thin body does not allow erase by body bias) S. M. Jung, et al., IEDM Tech. Dig. 2006 Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

6

3

25/02/2010

3D NAND layout

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

7

Reliability impact

• No degradation seen in 2nd layer • P/E characteristics not degraded either Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

8

4

25/02/2010

TFT 3D NAND

60 nm

90 nm

E. K. Lai, et al., IEDM Tech. Dig. 2006

• Alternative: poly-Si finFET BE-SONOS (Macronix, 2006) • polySi TFT: a-Si LPCVD + 600°C annealing • tSi = 60 nm, W G = 90 nm, LG = 200 nm Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

9

Reliability

• Reasonable reliability, no significant degradation in the top layer Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

10

5

25/02/2010

Impact of grain boundary (GB)

T.-H. Hsu, et al., IEDM Tech. Dig. 629, 2009

• tSi = 50 nm, WG = LG = 30 nm • Annealing of 600°C for 30 hours results in 200 nm average grain size  30 x 30 nm2 channels can be marginally affected Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

11

TFT vs. bulk devices • Large TFT devices are worse than bulk (GB effect) • Small TFT devices are similar to bulk, because of nearly zero GB • Bad P/E distributions due to GB statistics GB distance

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

12

6

25/02/2010

3D (stacked) NAND limits • One decoder for each layers  area overhead • Critical (and costly) litho steps to make minimum feature size are necessary for each layer • Epy-Si layer is a high temperature process  number of stacked layers is limited by thermal budget  A simpler, cheaper 3D architecture is needed

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

13

2D NAND architecture BL1

string1

DSL WL1 WL2 WL3 SSL

string2

• NAND is inherently 3D: DSL – BL coordinate  x WL1 WL2 – String coordinate  y WL3 SSL – WL coordinate  z • In planar NAND, BL and string along the same direction •  3D NAND is straightforward Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

14

7

25/02/2010

Bit-cost scalable (BiCS)

• BL (x) and upper SG (y) select a vertical string • Planar WL (z) select a single cell along the string • String area = 6F2, cell area = 6F2/N Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

15

Process sequence • String in three steps (lower SG plug + memory plug + upper SG plug) • Slit to reduce disturbs • Only one critical litho step = memory plug • CG formation:

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

16

8

25/02/2010

String

• Punch + plug  polysilicon for best hole filling • Undoped or n- doped  depletion mode vertical FETs to avoid separate n+/p+ doping Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

17

From 6F2 To 4F2

• BL have 2F pitch, USG have 3F pitch since they must surround the plug (2x0.5F) and be isolated (F) • 2-layer USG can achieve 2F pitch  4F2 string area Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

18

9

25/02/2010

Program/erase • Program: – select strings by USG+BL – High CG for e- injection from poly-channel • Erase: – High BL/SL, CG = 0 – BBT holes by GIDL at n+SL/LSG – Holes populate poly channel and raise body potential – Holes injection into charge traps Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

19

Program/erase characteristics

• P/E window improves for decreasing diameter, as a result of the larger E-field in the inner oxide for small diameter Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

20

10

25/02/2010

Macaroni channel

• Better control of body potential  better STS • Higher density of grain boundaries  better VT distribution Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

21

Reliability issues

• Retention to be improved (SONS dielectric stack, ONO oxide would be damaged by HF diluted etch used to improve poly-poly contact in the plug) • Limited scaling of macaroni FET • Charge migration in the trapping layer mitigated by channel length in the z direction  no limits Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

22

11

25/02/2010

Advantages over 3D NAND 3d nand layers =

bc =

BiCS layers =

bc =

1 2 1 4 22 nm 2 2 2 1 3 1,333333 0,666667 4 1 0,5 1 1 6 2 3 3 2 4 1,5 5 1,2 6 1 7 0,857143 8 0,75 9 0,666667 10 0,6 11 0,545455 12 45nm 0,5 13 0,461538 14 0,428571 15 0,4 16 0,375

Feb. 26, 2010

2 3 1,5 1 0,75 0,6 0,5 0,428571 0,375 0,333333 0,3 0,272727 0,25 0,230769 0,214286 0,2 0,1875

3 4 1,333333 1 0,666667 0,5 0,444444 0,333333 0,333333 0,25 3 2 1 0,666667 0,5 0,4 0,333333 0,285714 0,25 0,222222 0,2 0,181818 0,166667 0,153846 0,142857 0,133333 0,125

4 1,5 0,75 0,5 0,375 0,3 0,25 0,214286 0,1875 0,166667 0,15 0,136364 0,125 0,115385 0,107143 0,1 0,09375

• 3D NAND = 4F2 while BiCS = 6F2  3D NAND better for 1 layer only • 12 layer-BiCS is equal to 2 layers – 4bc 3D NAND • 12 layer-BiCS is equal to 1 layer – 2bc at half F (e.g. 22 nm instead of 45 nm)  when planar Flash will be no more scalable (e.g. 10 nm), BiCS will show up in a previous technology node (e.g. 50 nm)

D. Ielmini, "Non volatile memories" – 3

23

Cost comparison

• Cf is relatively large for BiCS (USG, BSG, punch and plug) • Cv is relatively large for 3DNAND (each layer must replicate DSL and SSL, drivers, etc.) • Yield degradation for 3DNAND (epi-Si deposition with high thermal budget) Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

24

12

25/02/2010

Pipe-shaped BiCS •

Some problems to be addressed: 1. Bad reliability because of SONS 2. Series resistance due to diffused source line 3. Uncontrollable LSG due to thermal budget • Pipe-shaped BiCS solves these issues

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

25

Reliability improvement

• Reliability improvement thanks to SONOS instead of SONS structure (poly is plugged all at once, no need for HF etch) Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

26

13

25/02/2010

Other issues and solutions • Other issues include: – Gate material is p+ polysilicon, no metal gate (e.g. W or TaN as in TANOS) can be integrated because of etching limitations – Erase: GIDL at VG < 0 needed  circuit overhead

• Those issues are solved by Terabit Cell Array Transistor (TCAT), the Samsung answer to BiCS

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

27

Terabit Cell Array Transistor (TCAT)

• Very similar to BiCS but: – Gate-last process where sacrificial nitride is replaced by metal gate (W) Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

28

14

25/02/2010

Gate-last process

Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

29

Program/erase characteristics

• P-doped channel connected to the substrate allows for bulk erase similar to NAND (not in BiCS with n-channel tied to n+ source) • Erase saturation at VT = -1 V to be improved by diameter scaling Feb. 26, 2010

D. Ielmini, "Non volatile memories" – 3

30

15

Suggest Documents