Hybrid Magneto-optic-silicon memory with an RLL code for data compression. (Technote).

The search for 3D (Three Dimensional) computer memories has intensified in the last year since the realization that 2D recording was approaching saturation.

Magneto-optical (MO) recording [1,2,3,4] is a particularly suitable technology since most MO alloys are amorphous and can be deposited in thin film form on single crystal silicon. This note reports on the feasibility of a novel 3D memory consisting of a high-speed silicon short-term cache memory and a long-term high density MO memory. This hybrid memory requires a suitable channel code. A RLL (run length limited) code [5] offers better dc stability, accurate timing synchronization and can support a higher storage density if a variable length code is used. RLL code is normally defined by code rules that can support a higher density into a larger sequence of channel bits during encoding. RLL (2,7) code has at least 2 'O's following each channel bit '1' and at most 7 'O's. In addition, the constant channel clock must correspond to a constant data rate. This note reports for the first time a "DAISY CHAIN" RLL code. In every case data compression occurs and the MO memory capacity is increased dramatically by as mu ch as 8 times. This "DAISY CHAIN" links each memory level to the nearest neighbour level and no wastage occurs at the memory edges. Some very small amount of error correction is required for major defects by comparing the MO recording with the original silicon memory from which it originated. Normal serial data word processing is used but as high quality multi-wavelength lasers and position sensitive detectors become available, parallel processing at very high speed for communications should be possible.

Excellent magneto-optic properties were obtained at Keele University by sputtering amorphous films of terbium-iron-cobalt onto single crystal silicon. Coercivity of close to 8,000 oersted on silicon substrates was very close to the value achieved on commercially available magneto-optic media on a polyearbonate substrate and almost identical to the polycarbonate substrate results at Keele University. [6] With a special overcoat these rare earth films can give up to a 20 times magneto-optic signal enhancement.

Since these films are vertically magnetized the flinging fields are very small and inductive readout is very difficult. These vertically magnetized regions form 'I's or 'O's depending on the external field direction, either up or down. Unmagnetised or longitudinal magnetized regions can be used as spacers to separate the code words. Magnetization of '1's and 'O's is carried out when the laser induced temperature is the Curie Temperature minus the ambient temperature. Erasure can be carried by using a heating cycle and a small vertical magnetic field to convert all the '1's to 'O's or vice versa.

In the hybrid memory on the encoding side of the digital channel a 16 bit digital data input word stream is replaced by a variable length (2,7) RLL code. Since the (2,7) code words are smaller than 16 bits they require less space so they can be processed at a faster speed.

To link the (2,7) code words the second is subtracted from the first when the '1's are synchronized, and the third from the second and so on. This 'daisy chain' of word differences means that for most applications error correction and synchronization will not be required with the magneto-optic part of the memory.

Because the silicon memory is limited to, say, 1 MB the subtracted (2,7) code words are stored on one of the three magneto-optic storage when the silicon memory is full. An example of 5 (2,7) subtracted code words for storage is given below. Note that MI is the origin and is stored first.

M1 0100

M2 1000

M3 001000

M4 000100

M5 0001000

M1 0100

M2-M1 0--0

M3-M2 00--

M4-M3 0--0

M5-M4 --0

If 16 bit (or 19) words had been used then 80 bits of magneto-optic memory would be, required. In practice, in the example shown, only 29 bits are required. These bits include one '1' and ten 'O's and 18 bits where there is no net magnetization. Consequently, only 11 laser pulses are required instead of 31 for the original five (2,7) code words. Laser pulses of 200ns in a 400 oersted bias field is typical for a Curie Temperature of about 200 degrees C. HENCE, RECORDING OR READING IS SPEEDED UP BY APPROXIMATELY 3 TIMES.

Since only 29 bits are required instead of 80 this gives an EXPANSION OF THE MEMORY of 4 x.

To summarize the encoding process:

1. INPUT 16 bit word digital data stream

2. Look-up table

3. (2,7) codeword in volts - plus 1 volt = '1', minus 1 volt 'O'

4. Voltage threshold detector set at 0.8 volts to detect '1's

5. Delay lines set at 23T, 19T, 13T, 6T for the above example

6. Shift registers

7. MI, M1 - M2 etc. stored initially in the silicon memory

8. MI, M1 - M2 etc. transferred to the magneto-optic memory

Note that decoding starts with MI as the origin then M2, M3 etc. can be derived and converted back to the original data using the look-up table.

Also note that an electromagnet must be used with the magneto-optic media to ensure that three states can be achieved - '1', 'O' and -- clock T space

Finally, Fibre Optic writing and reading with 0. 1 micron spot size has been achieved by narrowing a single mode 5 micron fibre.

Calculation of compression factor

Assuming that input words 16 bit long are used, then, since subtraction of successive words is required for the channel RLL code, two input words of a total length of 32 bits are required, Taking the example of Figure 1, for (2,7) code, there can be 5 'O's before the 1' and 7 'O's after the '1'

Hence:

'O's 1 to 12

unused 31 to 20

12 words and 306 bits unused. 6.5 bits on average out of 32 bits. Therefore, 20% are used and the compression is x5.

Conclusion

Data compression of up to x8 is achieved by forming a daisy chain of channel RLL variable length code words. The daisy chain is formed by subtracting the code word from the next code word neighbour in time for serial processing. The '1' in the code words are aligned so that synchronization is obtained. Error correction is minimized because the subtraction is carded out in a silicon memory which has zero defects in it. Any major defects in the MO memory can be registered and corrected for.

Further work on error correction and the calculation of the highest compression factor is in progress. Memories of 10. 8 TB on a credit card are now theoretically possible.

References

(1.) E.W.Willams, 'The CD-ROM and optical disc recording systems', Oxford University Press, Oxford, 99-13 8, (1996)

(2.) A.B.Marchant, 'Optical recording', Addison Wesley, Reading, 68-83, (1990)

(3.) M.Kaneko and A.Nakaoki, 'Recent progress in magnetically induced super-resolution', J.Magn. Soc. Jpn, Vol. 20, S 1, 7-12, (1996)

(4.) K.H.J.Busehow, 'Introduction to erasable MO recording', Nato ASI Series, Series E: Applied Sciences, Vol. 229 (1992)

(5.) K.E.Schoehamer Immink,' Coding techniques for digital recorders, Prentice Hall, I- 297, New York (1991)

(6.) E.W.Williams, to be published.

Acknowledgement

For the support of Mike Downey, Cavendish Management Resources and KHD Limited.