SCSI and ATA Disk Interface History: SASI, SCSI, IDE/ATA, Serial ATA and SAS (1979-2009)

Short answer: the host-to-disk interface moved the controller into the drive twice, once through SCSI and once through IDE/ATA, and then both moved from parallel ribbons to serial links. SCSI started as Shugart Associates' SASI, was taken up by the ANSI committee X3T9.2 at its April 1982 meeting and was approved as ANSI X3.131-1986 on 23 June 1986. ATA came from Compaq, Western Digital and Control Data's Imprimis division in 1986, was drafted in the CAM committee from 1989 and was approved as ANSI X3.221 on 12 May 1994. The serial successors were Serial ATA (working group announced February 2000, specification 1.0 dated 29 August 2001, version 1.0a in early 2003) and Serial Attached SCSI (INCITS 376, approved November 2003, announced by ANSI on 28 January 2004).

Before the device-side controller. The SCSI-1 foreword explains the problem the interfaces answered: controller logic had once been shared among as many devices as possible, but VLSI device controllers now made it possible to build one into each device, which it describes as particularly true for high-performance storage because recording-medium and read-channel problems are specific to each technology and are best resolved inside the device. The same foreword says that 'by 1982' the need for a device-independent peripheral bus was widely recognised in X3T9. On the PC side, PC Guide says the first disks with integrated controllers were the 'hardcards' of Quantum's Plus Development division, which put a 3.5-inch disk on a controller card in an ISA slot; this is one hobbyist-site account and is not confirmed by a second source here.

SASI at Shugart, about 1979-1981. In his Computer History Museum oral history, Larry Boucher says he inherited the job of building a disk controller for the eight-inch Shugart SA1000 as director of design services, that a colleague (Richard Steele) had already named his low-level device interface 'SASI' for 'Shugart Associates Storage Interface', and that Boucher asked to swap names, giving Steele 'SADI' (Shugart Associates' Device Interface), because he felt 'Shugart Associates System Interface' was the right name for his higher-level design. He says he split the design into a logical interface, a physical interface and a thin layer meant to ease a later move from parallel to serial, wrote the logical specification himself, had a company called DTC build the controller, and left Shugart in 1981 to start Adaptec. All of this is Boucher's own account; the SCSI-1 foreword confirms only that SASI was a commercial small-system parallel bus that 'had enjoyed significant market success' and was offered to X3T9.2 as the basis for a standard. The sources opened do not give a first-publication date for SASI (excerpts of a SASI design specification and a Usenet history gave dates for 1979 to 1981, but those pages were not opened, so no SASI date is stated here).

SASI becomes SCSI, 1982-1986. The foreword to ANSI X3.131-1986 says X3T9.2 chose the name Small Computer System Interface and began work at its April 1982 meeting, held plenary sessions every two months, and that the standard is 'a formalization and extension of the SASI'. It lists four extensions: a differential electrical option for cables up to 25 m, a synchronous transfer option allowing 3 to 4 megabytes per second, an optional extended command set with a block address space of 2 to the power 32 blocks against 2 to the power 21 for SASI (the powers are superscripts in the PDF and are not clear in its extracted text) plus self-identifying INQUIRY commands, and command sets for tape, printers, processors and optical disks. It records that at the February 1984 meeting representatives of Adaptec, Adaptive Data, Data Technology Corp., Fujitsu America, NCR and Shugart said for the record that their companies had implemented and tested SCSI. Boucher's recollection of the changes is different in emphasis: he names NCR's demand for larger addressing and Emulex's Bill Roberts's 'offset interlock' synchronous transfer as the two big changes, adds that the logical unit number field, which came from IBM 2314-style logical-versus-physical addressing and was nearly removed, survived, and says DTC wanted changes for single commands. On the name, he says the committee would not accept SASI, that he wished to keep it close to 'Small Computer Storage Interface', and that Dal Allan pronounced the new acronym 'scuzzy' while Boucher had hoped for 'sexy'; the pronunciation story is one participant's recollection. The standard itself spells the name 'Small Computer System Interface', so the expansion given in the oral history does not match it.

What SCSI-1 specified. ANSI X3.131-1986, approved 23 June 1986, defines a daisy-chained bus with both ends terminated and all signals common to all devices, with single-ended drivers (maximum cable 6 m, mainly inside a cabinet) or differential drivers (maximum 25 m, mainly outside), 18 bus signals, a data bus of 8 bits plus parity, a maximum of eight SCSI devices, multiple initiators and targets, and the 50-conductor connector. Its foreword states a goal of operating at up to 4 megabytes per second over up to 25 m. An abstract describes it as attaching small computers to 'low- to medium-performance' intelligent peripherals such as rigid disks, flexible disks, tape, printers and optical disks. The foreword also notes that Bernoulli-effect flexible disks and optical disks were already available with integral SCSI controllers (compare the removable cartridge card on this site).

The conflict over when NCR pushed for standardisation. Boucher says NCR had 'fallen in love' with SASI, that NCR's John Lohmeyer, already active in ANSI, took the idea to Shugart and to Boucher, and that this 'must have been in 87, late 87 or 88', after he had started Adaptec. The standard's foreword puts the start of X3T9.2's work in April 1982, and the SCSI-1 standard was approved in 1986, so Boucher's year cannot be right as stated; this card does not reconcile the two and does not use his year. He also describes a political reason for the 'low end' positioning: he says some committee members wanted IPI (a higher-end interface) and that SCSI was positioned as the low end to find a way into ANSI. That is one participant's account.

SCSI-2, 1985-1994. The introduction of the SCSI-2 working draft (revision 10L, 7 September 1993) says that when SCSI-1 was being finalised in 1985, several manufacturers wanted stronger mandatory requirements and more features for direct-access devices, so an ad hoc group wrote a working paper called the Common Command Set (CCS), and that SCSI-2 included the CCS results and extended them to all device types. It says SCSI-2 allows up to 16 devices on the bus, synchronous transfer up to 10 mega-transfers per second ('Fast SCSI'), a 32-bit wide option using a second cable with up to 40 megabytes per second, command queuing of up to 256 commands per initiator per logical unit, and that parity and message support became required; the new document is described as nearly three times larger than SCSI-1. T10 states that SCSI-2 is catalogued as X3.131:1994[1999] and that SCSI-1 was superseded by it. T10 also explains that the 'SCSI-3' name was dropped from later standards because it was cumbersome and that 'there is no SCSI-4 project envisioned'.

IDE and ATA, 1986-1989: Compaq, Western Digital, Imprimis. ATA-ATAPI.COM says the interface now called ATA was originally developed in 1986 by three companies: Control Data's Imprimis division, Western Digital (then mainly a silicon supplier whose controller chip IBM used in the PC/AT from 1984) and Compaq. It says the interface, called IDE by some and AT Attachment by others, was at first used only by Compaq, whose drives were mostly Conner Peripherals 3.5-inch models, and became popular when Conner sold drives in retail stores and other drive makers began making IDE/ATA drives. PC Guide independently names the same three companies and says Compaq began selling PCs with integrated hard disks and Western Digital controllers from the Deskpro 386 in 1986 (this is the only source here for the model). The ATA draft of 1994 says that in October 1988 a number of peripheral suppliers formed the Common Access Method (CAM) Committee, with the primary goal of a common software interface for SCSI peripherals and a secondary goal of specifying 'what is known as the AT Attachment interface'. ATA-ATAPI.COM gives 1 April 1989 as the date the first document describing the interface was submitted to the CAM committee; the OS/2 Museum lists 'DAD (Disk ATBus Definition) Rev 1, March 30, 1989' as the first, very incomplete, working draft, so the sources differ by two days and the card gives both.

ATA-1, 1989-1994. The foreword to the ATA draft says the standard was developed by X3T9.2 during 1986-90, with the approval process starting in 1991. The OS/2 Museum lists the surviving drafts (revision 2.1 of 11 June 1990, 2.2 of 15 August 1990 as 'the last under the CAM heading', 2.3 of 30 January 1991 under X3T9.2, 3.0 of 22 November 1991, 3.2 of 16 October 1992, an ANSI draft of 3 January 1993 and revision 4 of 17 March 1993) and says ANSI X3.221 was officially approved on 12 May 1994 with its text finalised in late 1993. Its author reads these drafts as showing that the 1991 versions had no LBA or removable-media support but did include IDENTIFY DRIVE, optional READ/WRITE MULTIPLE and SET MULTIPLE MODE, slave DMA and PIO modes 0, 1 and 2 (up to 8.33 megabytes per second), and that revision 3.2 of October 1992 is the oldest he has seen with LBA. The 1994 draft itself (revision 4c) defines ATA as 'a compatible register set and a 40-pin connector and its associated signals', gives a 16-bit data bus (DD0-15), two devices per cable (Drive 0 and Drive 1, called master and slave in industry usage), a cable length of 0.46 m (18 inches) unless both ends can be controlled, and lets a drive work in CHS or LBA mode command by command, with the LBA taken from bits 0 to 27 (28 bits). The draft's PIO timing table gives minimum cycle times of 600, 383 and 240 ns for modes 0, 1 and 2; at 16 bits per cycle those correspond to about 3.3, 5.2 and 8.3 megabytes per second (card calculation from the table; the 8.33 figure matches the OS/2 Museum). ATA-ATAPI.COM adds that ATA-1 was withdrawn as an ANSI standard in 1999 on T13's recommendation and that its author regards 'IDE' as the Conner, Compaq and Western Digital name from 1986-1987.

ATA-2 to ATA/ATAPI-7, 1996-2005 (single-source detail). Everything in this paragraph is from ATA-ATAPI.COM unless stated. ATA-2 (ANSI X3.279-1996) introduced PIO modes 3 and 4 and multiword DMA modes 1 and 2, for transfers up to about 16 megabytes per second; 'EIDE' was Western Digital's name for ATA-2 drives from 1993 or 1994, while 'FASTATA' was used by Seagate and Quantum, and the author calls all of them marketing names for ATA. ATA-3 (X3.298-1997) added SMART and Security but no new transfer modes. ATAPI, which carries SCSI-like packet commands over ATA so that CD-ROM and tape drives can share the cable, began as the SFF-8020 specification from a group of CD-ROM companies with help from Western Digital and Oak Technology, not as an ANSI standard. ATA/ATAPI-4 (INCITS 317-1998) added Ultra DMA with a CRC check at up to 33 megabytes per second and command queuing; ATA/ATAPI-5 (INCITS 340-2000) added Ultra DMA 66; ATA/ATAPI-6, whose work finished in October 2001, added Ultra DMA mode 5 (100) and raised the LBA width from 28 to 48 bits and the sector count from 8 to 16 bits; ATA/ATAPI-7, completed in early 2005, added Ultra DMA mode 6 (133) and the T13 version of Serial ATA. The 28-bit LBA limit of the original ATA corresponds to 2 to the power 28 sectors; at 512 bytes per sector that is about 137 gigabytes (card calculation, assuming 512-byte sectors).

Serial ATA, 2000-2009. Intel's press release of 21 November 2000 says the Serial ATA Working Group effort was announced at the Intel Developer Forum in February 2000, that Draft Specification 1.0 for 'Ultra SATA/1500' at 1.5 Gbps (scalable to 2x and 4x) was complete, and that the seven promoters were APT Technologies, Dell, IBM, Intel, Maxtor, Quantum and Seagate. SATA-IO's 20th anniversary page instead says SATA was introduced in February 2000 'through the efforts of APT Technologies, Dell, Intel, Maxtor, and Seagate'; the two promoter lists differ and the card shows both. The final revision 1.0 specification is dated 29 August 2001 and lists six promoters (APT, Dell, IBM, Intel, Maxtor, Seagate); InternetNews (30 August 2001) reports its release at the Intel Developer Forum and quotes the working group chair expecting the first Serial ATA/1500 storage products in the first half of 2002 (a vendor forecast). The specification's goals include being the 'primary inside-the-box storage connection', being 'completely SW transparent w/ ATA', a low pin count (2 pairs), a cable under 1 m, a transfer rate above the best parallel ATA of about 150 MB/s with scalability, and no jumpers or external terminators; the electrical layer uses 8b/10b encoding; connectors are hot-pluggable by staggered contacts and carry 12 V, 5 V and 3.3 V power; each drive has its own point-to-point cable and appears to software as a 'master' on a separate port.

SATA revisions. SATA-IO's archive says the 'Serial ATA final 1.0a specification was released on Feb 4, 2003', while the same organisation's anniversary table gives revision 1.0a as January 2003; this card shows both. The archive also lists the Port Selector 1.0 specification (28 August 2003), Extensions to Serial ATA 1.0a revision 1.2 (30 September 2004), Port Multiplier revision 1.2 (10 March 2005) and the integrated Serial ATA 2.5 specification (27 October 2005). The anniversary table gives revision 2.0 at 3 Gb/s in April 2004 and says SATA-IO was incorporated in 2004, revision 2.6 in February 2007 and revision 3.0 at 6 Gb/s in August 2008. SATA-IO's own press release announcing revision 3.0 is dated 27 May 2009 and says the specification offers up to six gigabits per second, adds an NCQ streaming command and NCQ management, and is backward compatible; so the release month of revision 3.0 is August 2008 in one SATA-IO page and May 2009 in another, and the card does not choose. Members' recollections on the anniversary page (single-source, not independently confirmed) describe early prototypes as an FPGA-based 'dongle' pair that converted parallel ATA to serial and back and ran existing operating systems without software changes.

Serial Attached SCSI, 2002-2004. The T10 project page for SAS (project 1562-D, INCITS 376, editor Rob Elliott) lists drafts from revision 00 of 12 May 2002 to revision 05 of 8 July 2003. INCITS announced on 6 November 2003 that the SAS standard, developed by T10 as INCITS 376, had been approved and quickly accepted as an American National Standard; the release calls it the next generation after parallel SCSI, says it defines rules for SCSI devices over a serial interconnect and also rules that let SAS hosts communicate with Serial ATA devices over the same interconnect, and quotes the T10 chair, John Lohmeyer, that 'all serial interfaces are self-clocking, so unlike parallel SCSI, data rates can be pushed higher and higher'. The SCSI Trade Association and INCITS then announced ANSI approval as ANSI/INCITS 376-2003 on 28 January 2004; their release (vendor claims) says the 3.0 Gb/s specifications were near-final about six months earlier, that several vendors demonstrated working products in the fourth quarter and were beginning to announce availability, that SAS preserves software investment in SCSI command sets, and that plugfests were planned for the year. A T10 proposal of 16 November 2004 for SAS 1.1 defines the link 'rate' with 1.5 Gbps and 3.0 Gbps as generation 1 (G1) and generation 2 (G2), and its reference list includes the Serial ATA II extensions and Port Selector documents; the T10 project page says the first SAS document was withdrawn in 2007 in favour of SAS-1.1. The sources opened here do not describe SAS 6 Gb/s or 12 Gb/s generations.

What the sources do not settle. The first publication date of SASI and its original capacity limits (the SASI specification itself was not opened); Boucher's date for NCR's standardisation push; the expansion of the SCSI acronym in Boucher's memory; whether the first ATA document was dated 30 March or 1 April 1989; the Deskpro 386 as Compaq's first integrated-controller PC (one source); the number and names of the original SATA promoters (five or seven in February 2000); the SATA 1.0a and 3.0 release months; and whether the 28-to-48-bit LBA change and the Ultra DMA mode details are stated as in ATA-ATAPI.COM, which is the only source read for them. Not covered: SCSI-3 and the Ultra and Ultra320 generations, Fibre Channel and SSA, ESDI, SMD, ST-506 and IPI, SATA after 2009, SAS-2 and later, and the move of SSDs to PCIe and NVMe.