Early Flash Memory and the First SSDs: Floating Gate, NOR and NAND, RAM Disks, SunDisk, CompactFlash, Fusion-io and Intel X25-M (1967-2008)

Short answer: flash storage began as a memory cell, not a disk. The floating-gate effect was described by Dawon Kahng and Simon Sze of Bell Labs in 1967; Fujio Masuoka of Toshiba turned a stacked-gate cell into 'flash', a one-transistor cell that is erased all at once, and presented the NOR form at the IEEE International Electron Devices Meeting (IEDM) in December 1984 and the NAND form at IEDM in 1987 (1988 in one source). Intel put NOR flash chips on sale in 1988. Disk-like flash products followed: a flash 'disk' board in 1988 to 1990 (one source), a 20 megabyte flash module for IBM from SunDisk in 1991, SunDisk's FlashDisk PC Cards from 1992, and the CompactFlash card announced in December 1994. Flash-based drives that looked like laptop and server disks arrived in 2006 to 2008, including Fusion-io's PCIe ioDrive (announced 25 September 2007) and Intel's X18-M and X25-M SATA drives (8 September 2008). Solid-state disks as a product, however, were older than flash: RAM-based units were sold from 1976, and for years they were the fast end of the market while flash was used for rugged, small or portable storage.

The floating gate, 1967 to 1980. The IEEE EDS brief says Kahng and Sze published the floating-gate memory effect in a Bell System Technical Journal brief; the text says they submitted it in May 1967 and published it in June 1967, while the brief's own reference list gives the August 1967 issue (volume 46, pages 1288 to 1295), so the card shows both months. Their first device had a 50 angstrom silicon dioxide gate oxide, a 1000 angstrom zirconium floating gate and an aluminium control gate, and needed pulses of 50 V and 0.5 microseconds to move charge onto the floating gate. Sze's stated motive was to find a semiconductor memory that would let a mainframe boot faster than it could from magnetic core. In 1971 Intel presented a 2 kilobit memory based on FAMOS (floating-gate avalanche-injection MOS), which had no control gate: programming took pulses of 50 V for 5 milliseconds, and erasing needed ultraviolet light, which the Semiconductor History Museum of Japan says took 20 to 30 minutes and a special package with a window that passes ultraviolet light. The same museum text says an electrically erasable device was announced by Yasuo Tarui and colleagues in 1971 but was held back by reliability, and that Intel's FLOTOX cell of 1980 made EEPROM practical at the cost of two transistors per bit. Masuoka, who joined Toshiba in 1971, says in his oral history that his double-gate EPROM was presented at the International Solid-State Circuits Conference (ISSCC) in 1973 and programmed in about 100 microseconds, with ultraviolet erasing as for Intel's parts; the EDS brief dates his stacked-gate SAMOS memory to IEDM 1972.

NOR flash, 1980 to 1985. IEEE Spectrum says Masuoka recruited four engineers to a semi-secret project in 1980; the transcript of his oral history garbles his own dating (he first says 1984 and then 1974 for the NOR idea, and 1982 for the work after a sales posting) and is not used to fix a date for the invention. The EDS brief quotes him that, by erasing all bits at once and then writing bit by bit, a select transistor was not needed, which he says could cut the cost per bit to a quarter, and that this was granted as US patent 4,437,174 (13 March 1984). Erasing was by field emission between the control gate and a third polysilicon erase gate. The cell was presented at IEDM in December 1984 using triple polysilicon, and a 256 kilobit chip at ISSCC 1985. The Semiconductor History Museum of Japan adds that writing was by Fowler-Nordheim tunnelling from the channel and that the name 'flash' came from the ability to erase all at once like a camera flash. IEEE Spectrum says Toshiba's managers told Masuoka to drop the idea and that Intel started its own NOR development after the 1984 paper, introducing a 256 kilobit chip in 1988. The 1994 Microprocessor Report article, by contrast, says Toshiba described the technology in 1985 but did not develop it into a commercial product until recently, and that Intel introduced its flash memories in 1988.

NAND flash, 1986 to 1996. Masuoka told the EDS brief that in 1986 he concluded NOR was not cheap enough to displace magnetic disks and that the answer was NAND, granted as US patent 5,245,566 (14 September 1993), with basic operation verified in April 1987 and presented at IEDM 1987. In his oral history he says he wrote about five NAND patents in 1986 while in Washington, D.C. on a Toshiba legal matter, that he developed the idea without permission, that he was refused funding for a 4 megabit chip by the computer side, and that he persuaded the head of Toshiba's consumer electronics development to fund it by proposing that a camera with 4 megabit NAND flash could replace film. The EDS brief gives 1988 for the working 4 megabit chip and ISSCC February 1989 for its announcement; IEEE Spectrum says Toshiba's first NAND flash reached the market in 1989. In the oral history Masuoka says the camera used eight chips, 32 megabits, was the first digital camera with NAND flash, and was made in 1990, and that the first chip had an endurance of about one thousand cycles, which improved with process work such as annealing. He says it took about four years from the idea to a good product, that his team never exceeded about ten people while he was there, that Samsung then proposed 16 megabit NAND with Toshiba, and that volume production of 16 megabit parts by both began from about 1996. Contemporary trade press disagreed with the later accounts on the state of NAND: the 1994 Microprocessor Report article says NAND was under development by Toshiba, National and Samsung 'but is not yet on the market', that Toshiba had run into problems in operation and yield and had to redesign the cell because high programming voltages reduced its durability, and that Toshiba was working with IBM on NAND-based PCMCIA cards. The same article describes Hitachi's AND cell and Mitsubishi's DINOR cell as competing NOR and NAND variants; neither is followed further here.

How the two cells differ, from the 1994 article. A flash cell is a field-effect transistor with an electrically isolated floating gate between the control gate and the channel. In Intel's ETOX NOR cell, programming used hot-electron injection and erasing used Fowler-Nordheim tunnelling to the source. The article says NOR cells connect in parallel to a bit line, give random access, and can be read in as little as 60 nanoseconds, but need a metal contact for every two cells; NAND cells connect in series and share one contact per string of 8 or 16 cells, which lets the cells be smaller, but reading a cell in the middle of a string means passing the others, so NAND read and write are slower, and the NAND cell uses tunnelling for both writing and erasing. It reports that writing a byte took as little as 4 microseconds and that erasing a block or a chip could take more than half a second, that the whole block must be erased to change one byte, and that endurance was typically 10,000 to 1,000,000 erase and write cycles. It also names the software answer: Microsoft's Flash File System spread wear over all blocks by periodically moving stable data and logging erase counts. The article's cost arguments are not repeated here.

Before flash: RAM, core and CCD 'disks', 1976 to the early 1990s. The earliest products called solid-state disks did not use flash. Dataram's Bulk Core, sold from 1976 for minicomputers from Digital Equipment and Data General, held up to eight 256K x 18 modules for 2 megabytes in one chassis (StorageSearch, PCWorld); PCWorld gives access times of 0.75 to 2 milliseconds. Whether it counts as semiconductor storage is disputed: PCWorld and StorageSearch describe RAM modules, while Jim Handy writes that it used core memory, and the 1977 Dataram annual-report scan that a search result surfaced could not be opened here. StorageTek's STC 4305 of 1978 emulated IBM's 2305 drum unit, stored 45 megabytes at up to 1.5 megabytes per second in charge-coupled devices (StorageSearch, PCWorld), and the Computer History Museum describes early commercial SSDs as volatile RAM chips backed by batteries and hard disks. StorageSearch quotes Fred Moore, then at StorageTek, that fewer than 25 CCD units were built before the company shifted to DRAM, with charge-coupled-device supplier problems a probable cause (his recollection, 2016). In the early 1980s RAM disks for personal computers followed: SemiDisk cards with 64 kilobit DRAM from 1982 (founder Jim Bell's recollection via StorageSearch) and Axlon's 320 kilobyte Ramdisk with a three-hour rechargeable battery (PCWorld). Two non-flash non-volatile attempts failed in the sources' accounts: silicon-nitride EAROMs of the late 1970s, whose 10-year retention and endurance fell short in the field and whose endurance rating was later cut to 1,000 cycles per cell (StorageSearch, from the editor's own memory of 1980), and Intel's bubble memory, positioned as a solid-state floppy disk and spun off in 1987 (see the magnetic bubble memory card). RAM-based units stayed in production through the 1990s as battery-backed SCSI boxes, such as NEC's of 1990 and Digital's EZ5x family of 1991 (StorageSearch).

The first flash disks, 1988 to 1994. PCWorld says a small Alabama vendor, Digipro, showed a prototype 'Flashdisk' board for IBM PC compatibles in 1988, made possible by Intel's NOR chips, that it held up to 16 megabytes, and that it shipped in January 1990 in 2, 4, 6 and 8 megabyte versions; it adds that M-Systems of Israel built a flash-drive prototype in 1989 but did not commercialise it until 1995. This is a single source and the Digipro claim is shown as such. SunDisk was founded in 1988 by Eli Harari, Sanjay Mehrotra and Jack Yuan (PCMag); the Computer History Museum says the company was co-founded to develop the technology for digital cameras and that in 1991 it built a prototype SSD module for IBM coupling a flash array with an intelligent controller that detects and corrects defective cells. StorageSearch instead says SunDisk 'shipped the world's first 2.5 inch flash SSD to IBM' in 1991 with 20 megabytes; the two sources disagree on prototype or product and neither names the IBM computer, so the claim that the drive went into a specific ThinkPad model is not made here. PCMag says SanDisk's PCMCIA FlashDisk line was first introduced in 1992 and was made until at least 2002. StorageSearch quotes Robin Harris's 1993 recollection of an HP Omnibook 300 option, a 10 megabyte flash PC Card (20 megabytes with compression) that he says lifted battery life from three or four hours to ten; the editor notes it was not strictly an SSD because it lacked wear leveling. By May 1994 the Microprocessor Report described SunDisk's 40 megabyte PCMCIA cards, built with patented technology from AT&T and flash chips made by Matsushita, with a custom controller designed with Motorola.

CompactFlash, 1994 to 1996. PCMag says SunDisk announced CompactFlash in December 1994, made it compatible with the parallel ATA interface used by hard drives, and showed four proposed capacities of 2, 4, 10 and 15 megabytes. The CompactFlash Association's FAQ also says CompactFlash was first introduced in 1994 and gives complete PCMCIA-ATA compatibility, a 50-pin connector against the 68 pins of PCMCIA, and card sizes of 43 mm by 36 mm by 3.3 mm (Type I) or 5 mm (Type II) (vendor claim for the rest of its characteristics). The Association says it was established in October 1995. Tech Monitor, reporting on 25 October 1995, says SanDisk had changed its name from SunDisk that summer after pressure from Sun Microsystems, and that a group of about a dozen companies headed by SanDisk, including Apple, Canon, Hewlett-Packard, Kodak, Motorola, Seagate and NEC, had established a proposed standard; it describes a 32 megabit postage-stamp-sized device that stored up to 15 megabytes uncompressed and a single-chip AT attachment controller, and it mentions the rival Minicard (Philips and Intel) and a Toshiba solid-state floppy disk card scheduled for January 1996. The Association's history page says CF Specification 1.0 was released in 1996 with an initial maximum capacity of 128 megabytes and support for up to 137 gigabytes, whereas PCMag says the original specification supported up to 128 gigabytes; these are shown as conflicting, and the 137 gigabyte figure agrees with the 28-bit ATA address limit calculated on the SCSI and ATA card. PCMag says Kodak's DC25 of 1996 was the first production digital camera to use CompactFlash, and that IBM's Microdrive of 1999 was a one-inch disk in a CF Type II shape.

Flash drives for computers, 2006 to 2008. StorageSearch and the Computer History Museum agree that in 2006 Samsung introduced the first high-volume Windows XP notebook using SSDs. StorageSearch lists 2007 as the year PCIe SSD shipments began in enterprise systems and gives 2005 for the founding of Fusion-io; the April 2008 Fusion-io release reproduced by Data Storage Connection says it was founded in 2006. StorageSearch reports that Fusion-io 'emerged from stealth mode' on 25 September 2007 at the DEMOFall conference with the ioDrive, a PCIe flash SSD claimed to reach up to 640 gigabytes, 100,000 I/O operations per second (IOPS) and sustained 800 megabytes per second read and 600 megabytes per second write (vendor claim), and said it would ship in the fourth quarter of 2007; the later release says the ioDrive 'will begin shipping on April 7th' and calls it the industry's first NAND-based enterprise storage solution (vendor claim), so the shipping date is shown as 'fourth quarter 2007 or April 2008'. On 8 September 2008 Intel announced that it had begun shipping the X18-M (1.8 inch) and X25-M (2.5 inch) SATA SSDs on multi-level-cell NAND, with the 80 gigabyte X25-M claimed at up to 250 megabytes per second read, 70 megabytes per second write and 85 microseconds read latency, a 10-channel architecture and its own controller and firmware to address write amplification and wear leveling (vendor claims), and said an SLC X25-E would follow within 90 days; Ars Technica, reporting the launch, says Intel announced its intention to enter the market about six months earlier and that its focus was on the performance and reliability of MLC drives, and repeats Intel's claims of better wear leveling and power efficiency without verifying them. StorageSearch says Intel began shipping the 32 gigabyte X25-E in October 2008 with 75 microsecond read latency and 250 and 170 megabytes per second read and write.

What the sources do not settle. The date of the NAND cell paper (1986, 1987 or 1988) and of Toshiba's first NAND product (1989 or later); whether Toshiba commercialised NOR flash before Intel; whether the Digipro Flashdisk shipped in 1990 as one source says; whether SunDisk's 1991 module was a prototype or a shipped product and what IBM did with it; whether the Bulk Core used core or semiconductor memory; the CompactFlash 1.0 capacity limit (128 megabytes with a 137 gigabyte address space, or 128 gigabytes); and the shipping date of the Fusion-io ioDrive. Not covered: Intel's own NOR history beyond the 1988 chip, AMD and Fujitsu, M-Systems DiskOnChip beyond a mention, USB flash drives, SmartMedia, MultiMediaCard and SD, the flash translation layer in any technical detail, multi-level cells before 2008, the 2008 SandForce and Indilinx controller generation, the NVMe protocol, and the move of SSDs into data centres after 2008.