Memory Device Evolution Cassino May 2008 Maurizio Di Zenzo Applications Lab Mgr
Agenda Random access memories A quick comparison of technologies Details of external memory technologies Solutions for low power and mobile applications Solutions for high performance Nonvolatile memories A quick comparison of technologies Details of external memory options 2
Random Access Memory Technologies (Relative Comparison) Technology High Speed SRAM Low Power SRAM Embedded DRAM External DRAM Pseudo Static RAM Power Density Area Speed Other Features Wide busses High density Medium density Costs 3
Architecture of a DRAM Memory device RAS_ CAS_ Timing and Control W_ OE_ Reference Voltages & Pumps Column Decode Column Decode A<11:0> Column Address Buffers (12) Sense Amps x8=1m 128k 128k 1Mx8=8M x8=1m 128k 128k Row Decode Sense Amps x8=1m 128k 128k 1Mx8=8M x8=1m 128k 128k Data In Reg (4) Column Decode 8M x 8 = 64M Column Decode I/O Buffers 4 out of 16 Selection DQ<3:0> Row Address Buffers (12) Sense Amps x8=1m 128k 128k 1Mx8=8M x8=1m 128k 128k Row Decode Sense Amps x8=1m 128k 128k 1Mx8=8M x8=1m 128k 128k Data Out Reg (4) 4
SDRAM Architecture DRAM Sense Amps Mid-Amp Output Buffer Column Decode SA 0 SA 1 SA 2 SA 3 SA 4 SA 5 SA 6 SA 7 30ns SDRAM Pipeline Queue 7ns Sense Amps Mid-Amp Output Buffer Column Address Counter & Decode CLK 5
Why Refresh? Bit Line (Column) DRAM Storage Cell Wordline (Row) Leakage Paths Charge leaks away over time. If nothing is done, in about a quarter of a second (200-300 milli-seconds), capacitors that were storing ones will be storing zeros! 6
Comparison of DRAM Timings RAS_ CAS_ ADDR DQ ROW COL 0 COL 1 COL 2 COL3 D0 D1 D2 D3 Enhanced page Mode RAS_ CAS_ EDO Mode ADDR ROW COL 0 COL 1 COL 2 COL3 DQ D0 D1 D2 D3 CLK RAS_ CAS_ ADDR ACTV ROW READ COL SDRAM DQ D0 D1 D2 D3 7
Basic DDR Differences from SDR Differential clock vs. single ended clock Differential clock (crossing) more accurate than rising edge 8
DDR DQS Strobe READ/WRITE READ WRITE 9
Ideal Signal Waveforms RAS_ CAS_ Note that an ideal signal is relatively smooth. 10
What is Real? Digital View Analog View 11
Basic DDR Differences from SDR Bi-directional data strobe (DQS) added to data bus Why a data strobe? Compensates for temperature, voltage, and loading Strobe travels with data During a WRITE to the DRAM, memory controller generates strobe Strobe occurs at beginning of data valid During a READ from the DRAM, the DRAM generates the strobe Strobe occurs at midpoint of data valid 12
Major differences from DDR to DDR3 Feature/Option DDR DDR2 DDR3 Package TSOP (66 pins) FBGA only FPGA 78-ball: x4, x8 FPGA 96-ball: x16 Voltage 2.5V 2.5V I/O 1.8V 1.8V I/O 1.5V Densities 128Mb 1Gb 256Mb 2Gb 512Mb 8Gb Internal Banks 4 4 and 8 8 Prefetch (MIN WRITE Burst) 2 4 8 Speed (Data Pin) 200 MHz, 266 MHz, 333 MHz, and 400 MHz 400 MHz, 533 MHz, and 667 MHz 800, 1600 Mb/s READ Latency 2, 2.5, 3 CLK CL + AL CL = 3, 4, 5 CL + AL CL = 3, 4, 5 WRITE Latency 1 clock READ latency 1 AL + CWL CWL = 5,6,7,8 13
DRAM Technology Trends 100% 90% Percent of Total Megabits Shipped 80% 70% 60% 50% 40% 30% 20% SDRAM DDR DDR2 DDR3 10% FP/EDO RDRAM 0% 2000 2001 2002 2003 2004 2005F 2006F 2007F 2008F 2009F RDRAM is a registered trademark of Rambus Inc. Source: Micron market research, isuppli 03/05 14
Main Memory Voltages 5 EDO 5V Projection 4 DRAM VDD (volts) 3 2 1 SDRAM 3.3V DDR 2.5V DDR II II 1.8V DDR III 1.5V 0 1995 1997 1999 2001 2003 2005 2007 2009 VDD is being reduced with each new technology I/O voltages are also being reduced Each decrease of VDD allows significant power reductions 15
Performance What Does This Mean? MB/s 5,000 4,500 4,000 3,500 3,000 2,500 2,000 1,500 1,000 500 0 64-bit Module MAX Bandwidth per Memory Technology PC2-4300 PC2-3200 PC3200 PC2700 PC2100 PC1600 EDP FPM PC100 PC133 DDR200 DDR266 DDR333 DDR400 DDR2-400 DDR2-533 16
Increasing DRAM Internal Prefetch 2500 SDRAM (1n) DDR SDRAM (2n) DDR2 SDRAM (4n) DDR3 SDRAM (8n) 2000 1600 1500 1333 1066 1000 800 533 667 667 500 267 333 400 0 67 67 100 100 200 200 133 167 133 167 167 133 167 200 1997 1998 1999 2000 2001 2002 2003F 2004F 2005F 2006F 2007F 2008F 2009F 2010F Internal Column Frequency (MHz) External Data Rate (Mb/s/pin) Source: Micron Marketing 09/03 17
Maximum Bandwidth Comparison Bandwidth per Data Pin (Mb/s) DDR2 800 BL4 DDR2 800 BL8 1,333 RLDRAM 533 SIO BL2 RLDRAM 533 SIO BL4, BL8 1,267 RLDRAM 533 CIO BL2 RLDRAM 533 CIO BL4 1,200 RLDRAM 533 CIO BL8 DDR3 1333 BL4 DDR3 1333 BL8 1,133 1,067 1,000 933 867 800 733 667 600 533 467 400 333 267 200 133 67 0 0.01 0.10 1.00 10.00 100.00 Read/Write Ratio 18
Pricing Comparison by Technology Relative Price 4 3 2 1 0 72Mb Price vs. Latency 0 10 20 30 40 50 60 70 16Mb 36Mb 18Mb 18Mb 9Mb Async SRAM QDR/DDR SRAM ZBT SRAM 4Mb 576Mb 288Mb RLDRAM 256Mb ns 1Gb DDR3 1Gb DDR2 512Mb DDR Speed 256Mb SDR 19
Technology Price/Mb Comparison Relative Price 40 30 QDR/DDR SRAM Price/Mb vs. Latency 20 10 ZBT SRAM Async SRAM RLDRAM Commodity DRAM 0 0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 ns 20
Low Power, Mobile Solutions
Mobile Features What Distinguishes it from Std SDRAM? Lower Power Targets top four power issues (Supply and I/O Voltage, Self Refresh Current and Operating Current) Typically 1.8V VDD and VDDQ option offered TCSR Temperature Compensated Self Refresh PASR Partial Array Self Refresh Deep Power Down Feature Variable Drive Strength VFBGA Packaging - Smaller Form Factor 22
Mobile Features Temperature Compensated Self Refresh First generation Mobile SDRAM designs allow manual setting of TCSR through the EMR (Extended Mode Register) All new (and some improved) designs incorporate an on-chip temperature sensor to dynamically control the self refresh rate automatically JEDEC LP-DDR specification had TCSR as optional if an on-chip temp sensor is used, the EMR bits must be a don t care to allow backward compatibility 23
Mobile Features Partial Array Self Refresh Ability to refrersh only one part of the total array. Programmed through the EMR (extended mode register) JEDEC LP-DDR specification had PASR as optional Full array, 1/2 array, 1/4 array, 1/8 array, 1/16 array Difficult to use this feature in the real application; most of the customers currently are not using it due to their inability to track where the data they require is exactly Most customers do request this feature to allow for future upgrades from their side 24
Mobile Features Deep Power Down JEDEC LP-DDR specification shows Deep Power Down as a design requirement Device must be in all banks idle state prior to entering DPD All memory data is lost DRAM must be fully re-initialized to ensure proper functionality LMR and EMR must be reloaded No customer usage currently known and no requests made to have this feature in future designs Micron will continue to offer this feature to be fully JEDEC compliant 25
Mobile Features Variable Output Drive Strength Offers customers the option to use a lower drive strength for lighter system loads or point-to-point environments Programmed through the EMR (extended mode register) standard full and half drive settings New series Mobile SDR/DDR SDRAM will include 1/4 and 1/8 drive setting in response to customer requests Customers are using this feature especially for the SDR parts that exhibit higher full drive strength than they required 26
Design Differences DLL Omitted from Design Reduces power consumption Minimizes the DRAM s ability to run at higher frequencies currently specified to 200MHz CL=3 and is pushing the limits of the design - customers pushing for tighter tac specifications No external VREF internal to the DRAM design Different initialization sequence required Offered in x32 configuration Main Specification Differences Lower IDD values Operating current and Self Refresh Some variations in timing Minimum speed specified on Std DDR, LP-DDR can run at very slow speeds Much higher tac (access time from CK/CK#) No DLL DDR Mobile SDRAM How Does it Differ From Std DDR? 27
Mobile Memory Comparison PSRAM LP-SDR LP-DDR LP-DDR2 (S4) MAX frequency 133 MHz 166 MHz CL = 3 200 MHz CL = 3 533 MHz CL = 8 Data rate (0% interleaving, 1:1 R/W, 1 Byte/access) 26.6 MB/s @ 133 MHz 16.6 MB/s @ 166 MHz 19.0 MB/s @ 200 MHz 16.7 MB/s @ 533 MHz Data rate (25% interleaving, 1:1 R/W, 1Byte-32Byte/access) Data rate (85% interleaving, 1:1 R/W, 32 Byte/access) Bus width (per device) 168.5 MB/s @ 133 MHz 212.8 MB/sec @ 133 MHz 120.7 MB/s @ 166 MHz (x16) 177.3 MB/s @ 166 MHz (x32) 167.2 MB/s @ 166 MHz (x16) 316.7 MB/s @ 166 MHz (x32) 269.0 MB/s @ 200 MHz (x16) 318.9 MB/s @ 200 MHz (x32) 684.5 MB/s @ 200 MHz (x16) 1196.3 MB/s @ 200 MHz (x32) 310.0 MB/s @ 533 MHz (x32) 2061.1 MB/s @ 533 MHz (x32) x16 x16, x32 x16, x32 x16, x32 Density 4 128Mb 64 512Mb 64Mb 1Gb 1 8Gb Banks 1 4 4 8 VDD/VDDQ 1.8V/3.3V 1.8V/1.8V 1.8V/1.8V 1.8V/1.2V* 1.2V/1.2V* 1.8V/1.2V/1.2V/1.2V** Package 54-ball VFBGA 60 ball VFBGA (x16) 54 ball VFBGA (x16) 90 ball VFBGA (x32) 90 ball VFBGA (x32) 152 ball VFBGA PoP* TBD JEDEC Functional pin Count (MAX Std (128Mb): 47 density, MAX width) AD MUX (64Mb): 31 (512Mb-x32): 57 (1Gb-x32): 62 (8Gb-x32***): 59 Burst options 4, 8, 16, 32* 1, 2, 4, 8 page 2, 4, 8, 16* 4, 8, 16 * Note: Not available for all densities; contact factory for availability ** Note: LP DDR2 has four supplies: VDD1/VDD2/VDDCA/VDDIO *** Note: Pin count does not change with density for LP DDR2 due to muxed cmd/addr bus 28
What is a PSRAM? Psuedo-SRAM s name comes from its similarity to SRAM; however, internally PSRAM is a DRAM combining the best of both products SRAM 6T cell architecture Broadside addressing (A0 Axx) Volatile hidden refresh Low power Low bandwidth Asynchronous access PSRAM 1T cell architecture Broadside addressing (A0 Axx) Volatile hidden refresh Low power Higher bandwidth Asynchronous access Synchronous access DRAM 1T cell architecture MUX d addressing (row col) Volatile refresh required High power High bandwidth Synchronous access 29
Architectural Comparison CellularRAM (PSRAM) LP SDRAM Addressing Broadside Multiplexed Refresh modes Hidden Auto/Self Row size 2k 4k (128K device) Interface optimized for? Random access Multibank access Bank architecture Single (1) bank Multiple (4) banks Density 4 128Mb 64 512Mb (SDR) 64Mb 1Gb (DDR) Voltage (Vcc/Vdd) 1.7 3.6V 1.8V Voltage (Vccq/Vddq) 1.7 3.6V 1.8V 30
Why Use PSRAM instead of SRAM? PSRAM can be used to replace sockets where SRAM has been traditionally used Feature Smaller die size (6T vs. 1T/2T) Benefit Lower cost part Increased density offering Ability to increase complexity and performance of end design Asynchronous interface Burst interface with variable latency FBGA and TSOP* package Drop in replacement Increased throughput and performance Small footprint for mobile applications * In design 31
Nonvolatile Technologies
Nonvolatile Technologies (Relative Comparison) Technology Power Density Speed read Speed write Endurance Reliability Availability Fuse (OTP) NA NOR NAND emmc MRAM PCRAM Requires block management Requires ECC and block management Protos 33
Flash Memory Cell Comparison NAND word line word line NOR bit line contact Cell array Unit Cell Unit Cell Layout Crosssection Cell size source line 2F 2F 4F 2 source line 2F 5F 10F 2 NAND Flash s small cell size enables high density and low cost 34
NAND vs. NOR: Basic Comparison NAND Advantages Fast writes Fast erases Disadvantages Slow random access NOR Advantages Random access Byte writes possible Disadvantages Slow writes Byte writes difficult Slow erase Applications File (disk) applications Any sequential data access Shadow for embedded applications Applications Replacement of EPROM Execute directly from nonvolatile memory 35
NAND vs. NOR: Strengths and Weaknesses Characteristic Random access read NAND Flash (SLC) MT29F4G08A 25µs (first byte) 25ns each for remaining 2,111 bytes NOR (Q Flash ) MT28F128J3 0.12µs Sustained read speed (sector basis) 33 MB/s (x8) 20.5 MB/s (x8) Random write speed ~220µs/2,112 bytes 180µs/32 bytes Sustained write speed (sector basis) 19 MB/s 0.178 MB/s Erase block size 128KB 128KB Erase time per block (TYP) 2ms 750ms NAND Flash is ideal for file storage, such as data or image files. If code is stored, it must be shadowed to RAM first (as in a PC). 36 NOR Flash is ideal for direct code execution (boot code).
MLC vs. SLC MLC SLC Features Bits per cell 2 1 Voltage 3.3V 3.3V, 1.8V Data width (bits) x8 x8, x16 Architecture Number of planes 2 1 or 2 Page size 2,112 4,314 bytes 2,112 bytes Pages per block 128 64 Reliability NOP (partial page programming) 1 4 ECC (per 512 bytes) 4+ 1 Endurance (ERASE/PROGRAM cycles) ~10K ~100K Array Operations tr (Max) 50µs 25µs tprog (TYP) 600 900µs 200 300µs tbers (TYP) 3ms 1.5 2ms Reliability of SLC is 10 times better! Performance of SLC ~3 times better! Cost premium is only about 2X Relative price 1X 2X 37
Another Option: emmc (Managed NAND) The next logical step in the high density NAND evolution for embedded applications Turns MLC NAND into a robust simple write/read memory Insulates the host from unnecessary details, including NAND block sizes, page sizes, planes, new features, process generation, MLC vs. SLC, wear leveling, and ECC requirements 38
What is Managed NAND (emmc)? Managed NAND combines the advantages of MLC NAND with a controller, which address all of the complexities of MLC (ECC, wear leveling, and block management) Provides MLC densities at close to MLC prices Advantages High density Lower cost (than SLC) MLC NAND Disadvantages Lower reliability (than SLC) Lower performance Managed NAND (emmc) Higher density (over SLC) Improved cost (over SLC) Improved reliability Higher performance (over MLC) Standard interface (MMC) Controller Advanced ECC Advanced wear leveling Advanced block management Advanced performance (dual channel) Single BGA package 39
What is emmc (Managed NAND) NAND Processor Block management ECC Driver NAND NAND bus emmc Processor Driver MMC Ctr ller ECC Block management NAND MMC bus Provides a consistent interface Published JEDEC /MMCA industry standard interface Addresses the increased complexities of current and future MLC flash devices Managed NAND Technology and vendor dependent Technology and vendor independent 40
Summary 41