PRML Read Channel (Partial-Response Maximum-Likelihood, 1970-2000)

Short answer: PRML is the read-channel method that replaced peak detection in hard drives. A drive head produces a weak, overlapping analog waveform; PRML equalizes it to a known partial-response shape, samples it, and uses the Viterbi algorithm to choose the bit sequence most likely to have produced the samples. The theory was published at IBM in 1970-71 by Hisashi Kobayashi, Ampex shipped the first product to use it (a tape recorder) in the 1980s, IBM shipped the first disk drive with a PRML channel, the 0681, in 1990, Quantum was the first independent drive maker to follow in 1993 (vendor claim), and the Computer History Museum says all hard-drive suppliers had adopted PRML by 2000.

1960s-1971: the theory. The pieces came from data communications. Kobayashi's 2009 history lists Lender's duobinary paper (1963), Kretzmer's 1966 generalization, Viterbi's 1967 paper on decoding convolutional codes and Forney's 1972 treatment of maximum-likelihood sequence estimation with intersymbol interference in its references. The abstract of Kobayashi and Tang (July 1970) says a magnetic recording channel can be regarded as a partial-response channel because the readback process differentiates, that conventional NRZI recording is equivalent to precoding that channel, and that the three-level output has redundancy that an error-detection scheme can use. The Eduard Rhein Foundation says Kobayashi's 1970 and 1971 papers pointed out that a recording channel is theoretically equivalent to a partial-response baseband transmission system and that the Viterbi algorithm can give maximum-likelihood detection of the coded sequence, with significant advantages over bit-by-bit decisions. Princeton's 2005 news item dates the conception and analysis to 1970 and 1971. A figure of 2.5 decibels of signal-to-noise gain is quoted for the 1971 paper in search-result summaries, but its abstract page could not be opened here, so the card does not use it.

Why nothing happened for years. The Rhein Foundation says Kobayashi's idea was ahead of its time: no one in industry or academia pursued it until, in the second half of the 1970s, Francois Dolivo at IBM's Zurich laboratory systematically studied signal-processing alternatives for disk drives, including a class IV partial-response read signal, and found significant potential gains in recording density. That is the foundation's account, written for an award, and no second source opened here describes the Zurich work in detail.

1983-1985: the first product was a tape recorder. The Ampex paper of June 1985 (Coleman, Lindholm, Petersen and Wood) describes a single recording channel at 117 megabits per second in which class IV partial response improves the effective signal-to-noise ratio and a Viterbi detector uses the redundancy of the ternary samples for further gain. The Computer History Museum says PRML was first applied in Ampex's 1984 recorder, which it calls the Digital Cartridge Recording System (DCRS), led by Charles Coleman, running at 28 megabits per square inch and 15 megabytes per second. Sources disagree on the details: StorageNewsletter says the Ampex DCRS shipped in 1983, the Wikipedia article says 1984 and calls it the Digital Cassette Recording System, and the Museum says Cartridge. The Wikipedia article adds that Ampex built an earlier prototype PRML channel at 20 megabits per second on an eight-inch disk drive but left the disk-drive business in 1985, and that Ampex's patent on the channel was never used; those two points rest on Wikipedia and its references only.

1986: the name. StorageNewsletter says the acronym PRML was invented by John Eggenberger of IBM San Jose around 1986, and that earlier names included partial response with Viterbi detection and class IV equalization with a maximum-likelihood detector; it cites US patent 4,707,681 by Eggenberger and Patel, filed 24 April 1986. This is a single secondary source and the patent was not opened here.

1990: the IBM 0681. StorageNewsletter says the 0681, code-named Redwing, was developed at IBM's Hursley laboratory in the UK from 1983 and first shipped in April 1990, in a 12-disk 857-megabyte formatted version (0681-1000) and an 8-disk 471-megabyte version (0681-500), both at 45.2 megabits per square inch (1,677 tracks per inch times 26,951 bits per inch). The Museum says the PRML algorithms were developed by a team led by Francois Dolivo at IBM Zurich and implemented in hardware at IBM Rochester, and gives the drive's figures as 45 megabits per square inch, 3 megabytes per second and a maximum of 471 megabytes. The two sources therefore disagree on the maximum capacity (857 against 471 megabytes); the card treats this as unresolved. The Wikipedia article gives the channel's data rate as 24 megabits per second, which equals 3 megabytes per second and so matches the Museum. The same article says the whole channel was one 68-pin PLCC chip and that an analog equalizer was followed by an adaptive digital equalizer; the Museum and the Rhein Foundation say the Rochester and Zurich groups worked together. The IBM paper of January 1992 (Cideciyan, Dolivo, Hermann, Hirt and Schott) describes this class of system: class IV partial response with maximum-likelihood sequence detection, a Viterbi detector simplified with a difference-metric algorithm, decision-directed gain and timing recovery, rate 8/9 constrained codes and an adaptive equalizer.

1993-1995: the second wave. Quantum's technical paper says IBM introduced a first-generation PRML channel in 1990 and that Quantum was the first non-captive drive maker to design one, first used in its Empire drives in 1993, with a second generation in the Grand Prix drives (vendor claim). Quantum describes the gain as a change of write code: a peak-detect channel using a 1,7 constraint stores two user bits for each three stored bits, while its PRML channel used a (0,4,4) constraint at 8 user bits for each 9 stored, and it reports 80,000 bits per inch for Empire against 63,600 for an Empire peak-detect model on similar heads and media, and internal data rates of 64 and 78 megabits per second against 48 (all vendor claims; 80,000 over 63,600 is 26 percent, as Quantum states). A second Quantum paper says it planned a third-generation single-chip PRML channel for the first quarter of 1995 and intended to combine PRML with magneto-resistive heads, saying PRML could raise areal density by 30-40 percent over peak detect. The Museum's 30-40 percent figure matches Quantum's, so the two are not independent confirmations of a measured number.

1996-2000: noise prediction. IBM Research's abstract of the 1996 ICC paper by Eleftheriou and Hirt describes noise-predictive maximum-likelihood (NPML) detection, which embeds a noise prediction and whitening process in the branch-metric computation of a Viterbi detector. It says simulations on Lorentzian channels showed substantial gains over both PRML and EPRML detectors and that NPML fits into existing PRML architectures. The Rhein Foundation says the observation behind it was that the noise accompanying the signal is not completely random, and that the new architecture entered IBM disk drive products in 2000; Princeton's 2005 item also says Eleftheriou's improvements were applied in IBM drives since 2000. Kobayashi's 2009 abstract says NPML and variants have been adopted by the hard-drive industry since 2000.

Recognition and what came after. In 2005 the Eduard Rhein Foundation gave its Technology Award to Kobayashi, Dolivo and Eleftheriou for this line of work (Rhein Foundation; Princeton). The Museum says modern channels, as of 2019, are about 10,000 times more complex in gate count than the 1990 channel and pass probability information back and forth between an NPML front end and an LDPC (low-density parity-check) back end on one chip. That LDPC step belongs to the later error-correction history and is not covered here.

What the sources do not settle. The 0681's maximum capacity (857 or 471 megabytes); the Ampex product's year and name; who named PRML (one secondary source); and how much of the density gain of the early 1990s was due to PRML as opposed to heads, media and servo. The 60 and 40 percent annual growth rates (Rhein Foundation) and the roughly 50 percent per year (Princeton) are award-text statements about the industry, not measurements tied to PRML alone. The Rhein Foundation, Princeton and Kobayashi's own 2009 article all come from or about the same group of IBM researchers, so they are not fully independent; the Museum and the Quantum papers add independent angles on the 1984, 1990 and 1993 products.