Anyone care to speculate on the limits of magnetic storage?

[jtr1962]

Any ideas on how large magnetic mechanical hard drives will get before some insurmountable technological limit is reached? It seems that every time we get close to reaching the limits of storage technology something new like IBM's "pixie dust" comes along.

I'll start with my thoughts on the matter. I feel "pixie dust" and more sensitive heads will get us to a few hundred GB per platter(standard 3.5" platters). At that point we will have reached the anisotropic limit of sputtered magnetic media. However, this will not be the end of the line. Nanomagnet research has suggested that by using patterned magnetic media it will be possible to reach areal densities a few hundred times greater than current drives, or on the order of a few TB per platter. Since track density is proportional to the square root of areal density, this suggests track densities of ~500,000 tpi. Servoing quickly and reliably onto such tiny tracks will undoubtably present a challenge, but one I feel will be overcome. Since patterned magnetic media will in essence be magnetic domains of a few thousand atoms, and I doubt that we will be able to servo on anything much denser than 1 million tpi except in a laboratory, I feel that several TB per platter will be the final limit of mechanical hard drives. Based on current progression, I would say this will occur in ~10 years(assuming a 60% areal density increase per year). Of course there will be far denser storage after that time, but it will not be magnetic. NV RAM and holographic are two technologies that come to mind. Whatever it is, however, I feel it will have no moving parts.

Any other thoughts or opinions on this?

 

[Tannin]

Well, JTR, I am hopelessly outside my field here. But seeing that has never stopped me in the past ....

Seems to me that this business of seeking accurately enough is the key problem. Obviously, we are getting an ever increasing requirement for a smaller "minimum adjustment size", so to speak - the head must be able to perform ever-finer micro movements in order to position itself exactly over a track. In itself this is no doubt a difficult problem. But it is compounded by the fact that the requirement for gross adjustments has not changed. We still need to be able to move the head an inch or so in 20-odd milliseconds, and even with a switch to smaller platters - 3.0 inch, 2.5 inch or in the extreme case 1 inch - the span of different sized movements that the actuator must be optimised for is getting bigger, not smaller.

Now (as will be obvious) I am no mechanical engineer, but it seems to me that we can think of this by analogy. Let's scale the thing up so we can visualise it.

Assume that you have been asked to make a radial arm (like a hard drive head or a record player stylus arm) and that this arm mechanism must traverse over the surface of a disc about the size of a large children's roundabout. (i.e., about 4 metres in diameter). Your original model (equivalent to an ST-225) must be able to pick out any of the 615 data tracks, and these tracks are spread across 1.5 metres. (4m diameter = 2m radius, allow a bit for the centre of the disc that is not used ~ say 1.5m.) Thus your arm positioning mechanism must be able to center on tracks that are 2.5mm apart. That's the starting point.

Now, we switch to a smaller playground roundabout (equivalent to a 3.5 inch disc) that has a 3 metre diameter. Let's model this one on a TM-262 which, like the ST-225, had 615 cylinders. Assuming a rough spaned distance of 1m this time, we get tracks 1.6mm apart. The game has got harder because the tracks are closer together, but the relative difference between a full seek and a one track seek has not changed.

Now we move to a modern drive. Let's see if I can drag some numbers up. ....

 

[Tannin]

Hmmm ... hard to find the appropriate figures. No-one seems to bother specifing the total number of physical tracks anymore. But I found some specs for the Deskstar 120GXP, which should be as good an example as any.

If I read IBM's figures right (they are not straightforward), a 120GXP has ~57,000 tracks (a quite astonishing number, by the way) which, when we substitute the numbers into our model, gives a single-track seek distance on our 3m roundabout of about 0.017mm. Picture that, gentlemen: you have an arm the tip of which travels through an arc of 1m, and you have to position it to centre over tracks which are 0.017mm apart! That's a tough ask. The arm is required to have an accuracy of better than 0.0002%!

Notice though, that this is "only" 100 times more precise than our TM-262 model roundabout, which required a seek accuracy in the order of 0.02%. I should imagine that the advance in linear density has been much greater in proportion than the advance in track density. Let's see if we can do some numbers for that too.....

 

[Tannin]

No! I was wrong. In fact (assuming I haven't gone and mislaid a zero or two) linear density has only improved by a factor of about 26! Check my sums, gentlemen: 17 SPT for a TM-262, 448 SPT for the inner track of a Deskstar 120GXP. Note that I'm using the inner track to provide a fair comparison - the TM-262 could not take advantage of different zones so the entire drive was limited to the abilities of the innermost track.

At first this surprised me - if my supposition is correct, that it would be more difficult to get more tracks per inch because of mecanical factors than it would be to get more linear data per inch (which is, I assume, mainly limited by magnetic and electronic issues). Notice, for example, that these new-fangled floppy drives that Panasonic have invented get 20MB out of an ordinary 1.44MB disc by retaining the 80 tracks but squeezing a great deal more data onto each one. They don't attempt to do, say, 360 tracks instead of 80.

But then I remembered that if moving the arm assembly accuurately enough is an issue today, how much more of an issue would it have been in 1985 when the TM-252 was a pup? Throw in the stepper mechanism instead of the voice coil, and it seems reasonable to retaim my hypothesis, at least for the time being.

 

[Tannin]

Right: enough reasoning. Let's see if we can't draw a conclusion or two from it. Right now, in our 120GXP model roundabout, we have an arm that is moving through an arc of 1m and has to be accurate to less than 0.02mm. If we are going to improve by a factor of 10 - no - let's make it a factor of 100, no point in messing about with relatively minor changes - then we need our arm to span that 1m arc with a precision of about 0.0002mm!

Seems to me that that is like trying to do some very fine micro-surgery with your elbow, wrist, and fingers in a cast. You just can't get that sort of accuracy out of your shoulder muscles!

The answer, if this analysis is correct, becomes obvious: the next step in seek mechanisms becomes an articulated head, one with, if you like, a "wrist" in it to take care of the fine adjustments.

Would this be mechanical? A tiny arm on the larger arm? Possibly, but I doubt it. A more practical sounding method would be some kind of flexible tip, perhaps made out of some you-beaut polymer that expands and contracts as you apply a voltage to it. Apply voltage to the left of the tip and it contracts and thus twists the head outwards, apply voltage to the right and it bends the other way. Now you have to stop it twisting in the horizontal plane. And deal with vibrations. Tricky stuff!

Or, maybe it would be all electronic. One can imagine an array of heads which can be switched between, or a very wide head which can be electronically "focussed" on any desired point across its width. ("Very wide", in this context, means perhaps 0.01mm - i.e. about 200 tracks.)

 

[jtr1962]

This might help your analogy a bit:

Product Specifications(Maxtor D540X-4G)

Number of Servo Sectors: 224
Data Zones per Surface: 16
Data Sectors per Track (ID/OD): 448/896
Areal Density (Gbits/in 2 max, ID/OD): 27.7/25.2
Flux Density (kfci, ID/OD): 506/461
Recording Density (kbpi, ID/OD): 486/442
Track Density (ktpi): 57


From the time of the ST-225, bit density has increased by about 40 times(I accounted for the smaller platter diameter of the Maxtor), while track density has went from 615 tracks/1.75"=350 tpi to 57,000 tpi, or a factor of 162! Apparently it has been easier to increase track density than bit density, but I think a fairer comparison would be between the first voice coil drives and current technology since the initial switch to voice coil actuators led to fairly large increase in track density. Going that route, I came up with the following:

ST-3051A


UNFORMATTED CAPACITY (MB) ________________N/A
FORMATTED CAPACITY (xx/17 SECTORS) (MB) __43.1
ACTUATOR TYPE ____________________________VOICE COIL
TRACKS ___________________________________
CYLINDERS __PHYSICAL/LOGICAL______________xx/820
HEADS ______PHYSICAL/LOGICAL______________ /6
DISCS (3.5 in)____________________________
MEDIA TYPE _______________________________THIN FILM
RECORDING METHOD _________________________RLL (2,7)
TRANSFER RATE INTERNAL (mbits/sec) _______up to 15
SPINDLE SPEED (RPM) ______________________3,211
AVERAGE LATENCY (mSEC) ___________________9.34
BUFFER ___________________________________32 KByte SeaCache
INTERFACE ________________________________AT BUS
SECTORS PER DRIVE ________________________84,270
TPI (TRACKS PER INCH) ____________________1,792

BPI (BITS PER INCH) ______________________31,005
AVERAGE ACCESS (ms) ______________________16
SINGLE TRACK SEEK (ms) ___________________3
MAX FULL SEEK (ms) _______________________28
MTBF (power-on hours) ____________________150,000
POWER REQUIREMENTS: +12V START-UP (amps) _0.9
POWER MANAGEMENT (Watts):
ACTIVE _______________5.3
IDLE _________________2.1
STANDBY ______________0.7
SLEEP ________________0.7
WRITE PRECOMP (cyl) ______________________N/A
REDUCED WRITE CURRENT (cyl) ______________N/A
LANDING ZONE (cyl) _______________________AUTO PARK
IBM AT DRIVE TYPE ________________________40, 44*



Seagate Technology Desk Reference Jan 11 2000

Track density has gone from 1,792 tpi to 57,000 tpi(a factor of 32), while bit density has gone from 31,005 bpi to 486,000 bpi, or a factor of 15.7). It seems that in both cases track density has gone up faster than bit density.

My theory about this is that servo technology is getting closer to approaching the ideal of being able to detect tracks spaced one bit apart, which would be 243,000 tpi in the case of the Maxtor 540DX. Of course, I'm assuming that you can have tracks with only one bit of space between them without them interfering with each other in some way. This assumption may be wrong. I would also imagine that the spacing is currently higher to tolerate servoing errors.

Interestingly, the new LS-240 can format a standard 1.44 MB floppy disk to 32 MB. Most of the gain is from the increased tpi that a servo rather than a blind stepper allows. If I remember correctly, the new format uses zone bit recording to get 36 to 53 spt instead of the usual 18, but 777 tracks instead of 80(thus taking greater advantage of the media's density in the radial direction).

 

[jtr1962]

The answer, if this analysis is correct, becomes obvious: the next step in seek mechanisms becomes an articulated head, one with, if you like, a "wrist" in it to take care of the fine adjustments.

Would this be mechanical? A tiny arm on the larger arm? Possibly, but I doubt it. A more practical sounding method would be some kind of flexible tip, perhaps made out of some you-beaut polymer that expands and contracts as you apply a voltage to it. Apply voltage to the left of the tip and it contracts and thus twists the head outwards, apply voltage to the right and it bends the other way. Now you have to stop it twisting in the horizontal plane. And deal with vibrations. Tricky stuff.

I've mentioned somewhere about the possibility of using nanotechnology for this. It is currently possible to make microscopic mechanisms(gears, motors, etc) etched directly out of a silicon wafer, so perhaps using the conventional voice coil for gross positioning(+- 100 tracks), and the micromachine for fine positioning. One advantage the micromachine would have would be a time constant on the order of microseconds instead of milliseconds. Another advantage would be that you could read multiple heads simultaneously since the heads on the head rack would no longer move exactly in unison(= increased STR). The net result of all this is that the drive will be very fast if seeks are confined to a 100 track area. If we assume 250,000 tpi, and a corresponding increase in bit density, then 100 tracks will be at least 100x2000spt, or 100 MB per head. On a 2-platter drive, this means that we can have seeks in the range of microseconds if we keep the partition to maybe 400 MB. This is currently enough to keep most operating systems on, so my idea can allow solid state like seek times(access times will actually be limited only by rotational latency) for the OS and standard(8 or 9 ms) seek times for the remainder of the drive just be making one small OS partition.