I find pretty interesting that the memory is not perfect out of the fab and its defects are coded on its label, hence these defects can be loaded to the device using it as a map, so it can work around these loops while working.
Now, this is even true for flash memory, but this is handled at the controller level and its dynamic during its initial "formatting" at the factory, so it's completely hidden from the host unless "host" is the controller.
OTOH, bubble memory comes with a "this machine has no brain, use your own" sticker on it.
Hard disks used to come with a paper label listing their bad sectors. (They still come with a list, it's just now it is hidden in the firmware like you describe.)
I do some retro-computing every now and then and of these old drives that still work (about a 50% chance), when I format them assuming all sectors are good, they will inevitably report the same bad sectors as is on the label.
I don't really know why "bad sectors are still bad after 40 years" is so surprising to me, but there you go.
There's a step which "flashes" the magnetic structure on the disks during manufacturing which also allows drives to navigate the disk surface and "see" individual sectors. The bad sectors list from factory shows the points where this process has failed due to bad coating or other anomalies.
So they are not worn out sectors, they were bad (or not there) from the start. You can't rub them out, because there is not enough magnetic flux or material to store data at that address.
I remember someone writing a rant against having the list in firmware and managed automatically by the drive. If the list was on the host side, your software could route around the bad sectors and minimize the total amount of seek time during a program run. But with the drive managing bad sector redirection, you could never predict read latency since any read could get re-routed to a spare sector by the drive, adding latency outside your control. Everything still became automatic though.
You could buy disk packs with no bad sectors back then if you wanted that ("bit error free"). They cost extra but the premium wasn't all that high. Of course they could still develop bad sectors later.
And the firmware remaps them, as far as I know. That's very different from a firmware exposing a bad sector list to the host, which is the case for some flash systems, I believe.
Disk firmware follows sectors which are hard to write, flags them as "current pending sector". Then reallocates that sectors data to one of the reserves when it has time and increases "Reallocated Sector Count", and if it deems the sector dead or unusable, it's get blacklisted and "Offline Uncorrectable" count is increased as well, from what I remember.
This is all transparent from the host, maybe an occasional "wait a little, I'm kinda busy" signal to the controller. Same for flash, but it's way faster and more transparent, since the real map inside the flash controller is never linear, anyway.
Not nearly all flash systems (including NAND based ones) use a dedicated controller. I believe e.g. iOS devices have for a long time still handled bad block handling in software, while macOS has always handled it in hardware.
Systems running macOS (i.e. Mac systems) use SATA or NVMe disks (depending on their age) which come with their own controllers, hence it's always isolated from the OS except SMART data. I have no idea about iOS systems.
Controller based management is also true for SD cards of any form, CF cards, USB flash drives or any external drives. This makes only embedded systems and devices like phones with direct NAND/NOR management AFAICS.
I worked at a place that evaluated bubble memory, but didn't use it (they had the bubble memory dev. boards). But they did use CCD memory. I remember it was prone to bit errors, which would have been OK if there had been any error correction, but there was not.
I really enjoyed this 2025 video from The Retro Collective explaining this technology (with some rather fun arts-and-crafts examples) and showing off a Konami arcade board (sadly nonfunctional, but still cool to look at) that used it.
The bubble memory was doomed by the fact that it had to be made from material that was patterned, by using photolithography or similar means.
Because of this, even if its structure was simpler than that of semiconductor memories, its bit density could not be increased much over that of semiconductor memories and its fabrication technology was not significantly cheaper.
Being much slower than random-access memory, while not having much cost advantage, made it non-competitive.
In contrast, the memories that use a homogeneous recording material with bits defined by the movement of a recording head, in the form of disks or tapes, compensate their slowness by having a much lower price per bit.
Thus bubble memories were too expensive in comparison with HDDs and too slow in comparison with DRAM, so no application remained for them in a computer that used DRAM and HDDs.
Which is why flash remains much more expensive than HDDs.
Flash is also much faster than a bubble memory, because it provides random access to blocks, even if a block is read sequentially. Thus flash was positioned in a good spot in a hierarchy of memories where a compromise is made between access speed and cost per bit.
To achieve a similar speed, a bubble memory would have to store a block, e.g. 4 kB, per bubble ring and have a huge number of bubble rings (i.e. billions, to compete with the current flash memories) to which random access would have to be provided by semiconductor ICs handling the multiplexing and demultiplexing of data.
Due to mixing ferromagnetic and semiconductor technologies, it is very unlikely that such a device could be made at an acceptable price.
Flash only got competitive after both HDDs and DRAM scaling slowed, allowing it to catch up, and end up in an interesting position between them.
There are about a dozen upcoming memory technologies with better theoretical headroom waiting in the wings for DRAM and Flash scaling to slow down. In the early 10's it really looked like Flash was about to be replaced, but then Samsung managed to move Flash to the third dimension. Very soon, either DRAM does the same, or it gets replaced by something.
Thanks for the in-depth explanation of this system. One of the most mainstream and famous use of bubble Memory may be the Gradius (Konami) shoot them up PCB and its infamous loading/warmup screen.
Robert Silverberg's novel "Shadrach In The Furnace", published in the 1970s, had a mention of some data being transported on a "billion-bit bubble chip". In retrospect, that's entertaining in 3 ways: 1) bubble memory didn't go anywhere, and 2) a billion bits is no longer an unthinkable amount of data, 3) solid state memory (which we'll take to include bubble chips here) still hasn't surpassed rotating memory (hard drives, tapes) at least in cost per bit. Maybe someday but not yet.
I had initially learned about bubble memory from the Logopolis serial of Doctor Who. I actually went and looked it up after watching it because the concept sounded neat
See also: https://en.wikipedia.org/wiki/Ferroelectric_RAM, which has been used in a few quite popular systems as well (e.g. the PS2 memory card supposedly uses it for some high-rewrite-count operations).
Very few of the MSP430s actually use FRAM, last I checked -- it's limited in both size and speed. (My understanding is that the FRAM speed is substantially better than Flash speed, but they treat it like SRAM for XIP, with I think two cycle waits up to 25 MHz or something, vs Flash where the expectations for XIP are much lower, so higher clock speeds are supported with more wait states.) The family is split between Flash (the majority) and the FRAM parts.
That reminds me of core memory, which is also nonvolatile and for a long time was the primary memory of many computer systems.
I suspect "growing up" with a system where the computer itself is nonvolatile (turn it off, move it, turn it on, it's doing exactly what it was doing when it was turned off) changed the way people looked at computers vs modern systems.
It's too bad intel didn't push harder on their optane stuff. I guess it's too bad (for intel) for lots of different abandoned / failed projects.
Nonvolatile memory has advantages, but also disadvantages.
In a computer that would have only non-volatile memory it would be difficult to ensure that things that must be secret remain secret.
Various events like power failures or software bugs could prevent the erasure of the secrets when they are no longer required.
Total memory encryption, like first introduced in server CPUs by AMD, and then followed by Intel, could solve this problem, but even that depends on the CPU having internal memory that is volatile, so this allows only the reduction of the amount of volatile memory that is needed, not its complete elimination.
I find pretty interesting that the memory is not perfect out of the fab and its defects are coded on its label, hence these defects can be loaded to the device using it as a map, so it can work around these loops while working.
Now, this is even true for flash memory, but this is handled at the controller level and its dynamic during its initial "formatting" at the factory, so it's completely hidden from the host unless "host" is the controller.
OTOH, bubble memory comes with a "this machine has no brain, use your own" sticker on it.
Hard disks used to come with a paper label listing their bad sectors. (They still come with a list, it's just now it is hidden in the firmware like you describe.)
I do some retro-computing every now and then and of these old drives that still work (about a 50% chance), when I format them assuming all sectors are good, they will inevitably report the same bad sectors as is on the label.
I don't really know why "bad sectors are still bad after 40 years" is so surprising to me, but there you go.
There's a step which "flashes" the magnetic structure on the disks during manufacturing which also allows drives to navigate the disk surface and "see" individual sectors. The bad sectors list from factory shows the points where this process has failed due to bad coating or other anomalies.
So they are not worn out sectors, they were bad (or not there) from the start. You can't rub them out, because there is not enough magnetic flux or material to store data at that address.
I remember someone writing a rant against having the list in firmware and managed automatically by the drive. If the list was on the host side, your software could route around the bad sectors and minimize the total amount of seek time during a program run. But with the drive managing bad sector redirection, you could never predict read latency since any read could get re-routed to a spare sector by the drive, adding latency outside your control. Everything still became automatic though.
You could buy disk packs with no bad sectors back then if you wanted that ("bit error free"). They cost extra but the premium wasn't all that high. Of course they could still develop bad sectors later.
And the firmware remaps them, as far as I know. That's very different from a firmware exposing a bad sector list to the host, which is the case for some flash systems, I believe.
Disk firmware follows sectors which are hard to write, flags them as "current pending sector". Then reallocates that sectors data to one of the reserves when it has time and increases "Reallocated Sector Count", and if it deems the sector dead or unusable, it's get blacklisted and "Offline Uncorrectable" count is increased as well, from what I remember.
This is all transparent from the host, maybe an occasional "wait a little, I'm kinda busy" signal to the controller. Same for flash, but it's way faster and more transparent, since the real map inside the flash controller is never linear, anyway.
Not nearly all flash systems (including NAND based ones) use a dedicated controller. I believe e.g. iOS devices have for a long time still handled bad block handling in software, while macOS has always handled it in hardware.
Systems running macOS (i.e. Mac systems) use SATA or NVMe disks (depending on their age) which come with their own controllers, hence it's always isolated from the OS except SMART data. I have no idea about iOS systems.
Controller based management is also true for SD cards of any form, CF cards, USB flash drives or any external drives. This makes only embedded systems and devices like phones with direct NAND/NOR management AFAICS.
I worked at a place that evaluated bubble memory, but didn't use it (they had the bubble memory dev. boards). But they did use CCD memory. I remember it was prone to bit errors, which would have been OK if there had been any error correction, but there was not.
https://deramp.com/downloads/mfe_archive/050-Component%20Spe...
In Patriot Games by Tom Clancy, there is a scene early in the book where Jack Ryan is talking about the laptop he is using to write his book.
I don't remember the exact specs but am 100% sure he mentions that it has bubble memory.
I just read the book and it does. I had to look it up when I read it because this was before my time.
I really enjoyed this 2025 video from The Retro Collective explaining this technology (with some rather fun arts-and-crafts examples) and showing off a Konami arcade board (sadly nonfunctional, but still cool to look at) that used it.
https://www.youtube.com/watch?v=DVJ9ED8mWt0
I remember a lot of talk about how transformative bubble memory was going to be circa 1980 and then quit hearing about it.
Indeed.
The bubble memory was doomed by the fact that it had to be made from material that was patterned, by using photolithography or similar means.
Because of this, even if its structure was simpler than that of semiconductor memories, its bit density could not be increased much over that of semiconductor memories and its fabrication technology was not significantly cheaper.
Being much slower than random-access memory, while not having much cost advantage, made it non-competitive.
In contrast, the memories that use a homogeneous recording material with bits defined by the movement of a recording head, in the form of disks or tapes, compensate their slowness by having a much lower price per bit.
Thus bubble memories were too expensive in comparison with HDDs and too slow in comparison with DRAM, so no application remained for them in a computer that used DRAM and HDDs.
Flash is patterned. There is talk about patterned surfaces for HDD but who knows if it will ever see the light of day.
Which is why flash remains much more expensive than HDDs.
Flash is also much faster than a bubble memory, because it provides random access to blocks, even if a block is read sequentially. Thus flash was positioned in a good spot in a hierarchy of memories where a compromise is made between access speed and cost per bit.
To achieve a similar speed, a bubble memory would have to store a block, e.g. 4 kB, per bubble ring and have a huge number of bubble rings (i.e. billions, to compete with the current flash memories) to which random access would have to be provided by semiconductor ICs handling the multiplexing and demultiplexing of data.
Due to mixing ferromagnetic and semiconductor technologies, it is very unlikely that such a device could be made at an acceptable price.
Flash only got competitive after both HDDs and DRAM scaling slowed, allowing it to catch up, and end up in an interesting position between them.
There are about a dozen upcoming memory technologies with better theoretical headroom waiting in the wings for DRAM and Flash scaling to slow down. In the early 10's it really looked like Flash was about to be replaced, but then Samsung managed to move Flash to the third dimension. Very soon, either DRAM does the same, or it gets replaced by something.
Thanks for the in-depth explanation of this system. One of the most mainstream and famous use of bubble Memory may be the Gradius (Konami) shoot them up PCB and its infamous loading/warmup screen.
Robert Silverberg's novel "Shadrach In The Furnace", published in the 1970s, had a mention of some data being transported on a "billion-bit bubble chip". In retrospect, that's entertaining in 3 ways: 1) bubble memory didn't go anywhere, and 2) a billion bits is no longer an unthinkable amount of data, 3) solid state memory (which we'll take to include bubble chips here) still hasn't surpassed rotating memory (hard drives, tapes) at least in cost per bit. Maybe someday but not yet.
I had initially learned about bubble memory from the Logopolis serial of Doctor Who. I actually went and looked it up after watching it because the concept sounded neat
https://tardis.fandom.com/wiki/Logopolis_(TV_story)
We used Intel 7110 magnetic-bubble memory in our mobile robot. CMU early 80s. VMEbus I think it was.
See also: https://en.wikipedia.org/wiki/Ferroelectric_RAM, which has been used in a few quite popular systems as well (e.g. the PS2 memory card supposedly uses it for some high-rewrite-count operations).
MRAM (magnetic and non-volatile) had some applications in GBA cartridges to replace small amounts of battery-backed SRAM.
FRAM is also used in all modern MSP430 microntrollers from Texas Instruments. Amazing technology
Very few of the MSP430s actually use FRAM, last I checked -- it's limited in both size and speed. (My understanding is that the FRAM speed is substantially better than Flash speed, but they treat it like SRAM for XIP, with I think two cycle waits up to 25 MHz or something, vs Flash where the expectations for XIP are much lower, so higher clock speeds are supported with more wait states.) The family is split between Flash (the majority) and the FRAM parts.
That reminds me of core memory, which is also nonvolatile and for a long time was the primary memory of many computer systems.
I suspect "growing up" with a system where the computer itself is nonvolatile (turn it off, move it, turn it on, it's doing exactly what it was doing when it was turned off) changed the way people looked at computers vs modern systems.
It's too bad intel didn't push harder on their optane stuff. I guess it's too bad (for intel) for lots of different abandoned / failed projects.
Nonvolatile memory has advantages, but also disadvantages.
In a computer that would have only non-volatile memory it would be difficult to ensure that things that must be secret remain secret.
Various events like power failures or software bugs could prevent the erasure of the secrets when they are no longer required.
Total memory encryption, like first introduced in server CPUs by AMD, and then followed by Intel, could solve this problem, but even that depends on the CPU having internal memory that is volatile, so this allows only the reduction of the amount of volatile memory that is needed, not its complete elimination.
[flagged]