I don’t get why storage isn’t sexier compared to CPUs. Vast amounts of cheap and fast storage is the enabler of our current computing paradigm, every bit as much as CPU tech. And it’s just as cutting edge! NAND is working right up against the limits of physics just like every semiconductor technology. There are hugely interesting problems being solved around optimizing cost and performance and capacity and scalability.
But all the buzz seems to be about RISC-V and Apple’s chips, and Intel’s woes. All the hardware superstars are CPU people. Who’s the Jim Keller of storage?
If you're looking for the sexy end of storage, send yourself down the rabbit hole of novel/upcoming NVM technologies. There is a wealth of interesting stuff to read and discover about MRAMs, CBRAMs, PCMs FeRAMs, FeFET, Racetrack memory. For more speculative/further away stuff you can look at DNA digital storage, quantum memory, time crystal memory, millipede rams and holographic storage.
There are some decent review papers and conference talks that cover the main contenders for flash storage replacement and their general benefits/drawbacks as well as some very dense research papers about the minutiae of each if you want to go deeper.
You're absolutely right about hard disks like this being essential to our current computing capability.
But in the personal computer space, very few people need lots of storage, so they will not "feel" any difference between 2TB and 20TB of hard disk space.
Even between SSDs, while some are undoubtedly dramatically better than others, for the average person they wouldn't necessarily notice that their notebook is ever so slightly faster to boot up or load their browser.
More advanced CPUs tend to have a much bigger effect on their day to day computing as their computer will feel faster and be able to run longer on batteries.
I think if huge amounts of storage were cheap and common, what people appear to “need” would naturally evolve to suit those capabilities. Imagine if you could click a button and make a copy of every major studio movie that’s been released in the last 50 years as easily as you copy an email.
I think personal space storage requirements have increased but consumers have been pushed towards cloud storage. People are drawn in by the free photo backup and when the free tier runs out, it's only $2/mo to continue making it someone else's problem. Business loves the recurring revenue stream and consumers love low/no upfront costs. As long as they don't get randomly flagged for a TOS violation and get banned with no recourse, it's fine not having local copies of their photos.
I don’t think it’s “very few”. NAS devices seem to be so common these days. Everyone is snapping photos and videos with their phones. Latest phones can record video in 4k. Then you have video games where modern ones eat up more than 100GB per installation. A lot of people need more storage. Not to mention professional photographers, I happen to know a few and they buy at least 2 4TB disks yearly to store everything in copies.
Apple's chips include storage. They put super fast storage in the SoC.
Raw capacity doesn't matter to most people. I collected up all of my personal files including all photos/vidoes/documents/etc, and it came out to about 35GB. The average person has absolutely no use for a 20TB drive so it isn't very interesting. For CCTV uses, 20TB drives sound awesome.
Fast storage is all that matters to regular users and not capacity.
I just looked this up to check and you are kinda right, It is a separate chip but it sits under the same heat spreader and on the same little pcb platform so I guess on some level they are separate but once they get joined they essentially form one part for most practical purposes.
>But all the buzz seems to be about RISC-V and Apple’s chips, and Intel’s woes. All the hardware superstars are CPU people. Who’s the Jim Keller of storage?
NAND and Storage technical improvement are relatively slow. And they also happens to be commodity, which means new technology improvement aren't as attractive if it doesn't bring down cost. A lot of the improvements happens in Operation and Scale.
Western Digital is a big supporter of RISC-V because they (probably) plan to replace the Marvell MIPS chips in the HDD devices with RISC-V.
But the reason you probably don't get storage superstars is because it's a healthy market. There are only a handful of cpu manufacturers/designers (amd, intel, Nvidia, apple, Marvell, ARM, Qualcomm) and because the instruction sets are not compatible it's a huge undertaking to move to a new architecture. All applications must migrate.
But with storage the application uses POSIX and you can out any hardware behind it. Or object API. Or a db. But mostly the application doesn't care except in very precise cases where performance matters a great deal. And even if one application needs object and another is ok with a posix API, it's ok because you can use both from the same machine. So the choices aren't as dramatic.
I can see the attraction for enterprise-level storage, where a single chassis might contain anywhere from 20-90 drives, but my thought on seeing news like this is that surely we're getting to the point where resilvering times will make small arrays infeasible from a reliability standpoint.
My home NAS only has four drives in it: if one 20TB drive dies, how long will I have to run on three, even if I have a spare onhand? Will they all survive long enough to rebuild? Or am I going to have to run some RAID-6/RAID-Z2-like array just to ensure durability?
Standard reminder to home / small biz RAID users: it is an availability solution, not a replacement for backups. It lets you have fewer service outages connected to disk failures. There are many more ways to lose your data than physical media failures.
This is for enterprise level storage. As you've said the resilvering time is too long unless you're using something like object storage with a whole cabinet to eat the reduced EC shard availability. But the good news is that you have access to enterprise level storage for your data that you want to make sure is durable. Use cloud storage.
it's still true that once you lose a drive you want to pop in a new one and resilver, it just reduces the risk to have 2 spares/parity drives instead of 1.
my array is relatively small and it still takes like 24 hours to resilver one drive. pretty brutal.
my current risk mitigation strategy is to have one spare in the array, and then have one 'warm spare' in the nas that is not part of the array and is instead a target of scheduled backups of the whole array every week. that way if i somehow lose another drive in the array while resilvering is in process i'm not totally screwed. (if my NAS could hold more drives I could do both, but more bays is way more price...)
naturally I also backup to the cloud. but you gotta have a local solution too.
Your usage is different than mine: I have a 18TB disk in my desktop tower so all my data is always available and have a second tower with 10x2TB disks in raid 6 for backups (all my old disks).
I switch on the second tower once a week and rsync everything from my desktop.
I don't have a solution if both fail at the same time but the probability of it happening is low enough that I accept the risk.
Ideally I would need a third backup rig somewhere outside my house, but I can't afford this level of safety.
It’s going to be difficult for me to regain trust in WD following recent cases where they swapped out controllers on consumer SSD products with components offering degraded performance following their initial release.
I bought nine of them and they are still in service, although they have always been slow and I wondered why until I got the class action settlement letter. I'm not even going to bother redeeming the settlement. It's something like $5 per drive.
I would love to be wrong. But nothing I see in the current investor as well as technology roadmap suggest we are going to see significant cheaper NAND production cost. ( Note: Cost per bit aren't dropping much with more layers incase anyone is wondering.)
I would be surprised if we could even push 4TB below $200 by 2025. Compared to a 4TB 2.5" external HDD which cost less than $100 today.
> Western Digital's 20TB HDDs with OptiNAND rely on the company's familiar nine-platter 7200RPM helium-filled platform that uses energy-assisted perpendicular magnetic recording technology (ePMR).
I was surprised when (a few months ago) I bought a few 14TB WD external drives, shucked one, and found a helium-filled drive.
people who are home building freenas/truenas type setups with like, 16 x 14TB drives have had generally really good results for shucking WD USB3 external drives for what's inside. You occasionally need to put kapton tape over the 3.3V pin, which is easy enough for a $5 roll of tape and a precision razor knife to cut it narrow enough.
If you're using the drives with normal sata power plugs (not in an enclosure), you can also just snip the orange wire. (And maybe electrical tape over the live end). IMHO, this is easier than precision cutting and applying tape... But it doesn't work if you've got an enclosure that supplies 3.3V.
When you have a power supply that's not SATA 3.3 compliant and a drive that supports it. The SATA 3.3 revision revised one of the 3.3v pins to be a power disable pin. Almost no hard drives support it so it's not really something you'll run into but consumer external drives will often be ordinarily super expensive enterprise drives in a plastic shell. Google the SKU of whatever external drive you're thinking of shucking to find out if it needs a SATA 3.3 power connector and instructions on opening it without damaging it in case you ever need to warranty the drive.
Having shucked numerous WD external drives - I can say I'm unsurprised. They have a history of putting decent off the shelf drives in their externals - once I even found a Black drive in one!
That's pretty much what they are. Hard drives are binned based on the yield of each production run. Server drives get the best lots, then tier one OEMs, etc. External drives are the lowest bin before the trash bin.
I'm actually not sure that's true. I run a NAS with 12 8TB drives, and not a single one has failed over several years. Nothing in SMART log either. All "shucked" cheap WD My Books. Running raidz-2. Sure, an anecdote, but how do you even "bin" something like this? It's not like you can vary voltage or frequency or fuse out busted cores (which is how they bin the CPUs).
It makes no mention of the physical layer, but spinning disks need to
1. increase their device bandwidth SAS only does what, 1200MB/s ? Networks are at 200-400G right now. And an individual SAS device can only do 12Gbits.
2. take predicates, even if the interface can handle 4-16GB/s your internal bandwidth and the ability to take a predicate means the drive can start understanding the data it stores
3. HDD physical media bandwidth needs to increase. Hard drives should move to 1-4U sizes, spin slower, have massively larger platters and multiple read heads.
If we assume 200GB/cm^2, we come out to ballparkish numbers for modern drives if we assume 5 platters. Lets just say that is true.
A single side of a 30cm platter is equivalent to 5 platters, two sides so 10 surfaces of a 9cm platter. Lets ignore the space taken up by the hub.
So a 19inch rack sized storage device is going to have 10x the storage capacity. It might have 10x the HD heads? Is more a chicken all in a single egg and ... I think would have to do something more impressive to get economies of scale.
The current 3.5" drive is going to pack a lot more than 10x to the 1u. So I think the current target I'd like to see drives go is a higher bandwidth access to the device, larger buffers, ability to send search predicates and a higher level protocol. We need HDD with NVMe. One should be able to do 4GB/s to an HDD for at least 10 seconds. All HDDs should be hybrid except for DVR drives as they only see there sustained write rate.
Hard drives are still first and foremost optimized for low cost per TB. Almost everything you're asking for would work against that goal, even if you could get economies of scale on these specialized drives. You would end up with little or no cost savings relative to QLC NAND, sequential read performance and random IO performance that would still be worse than QLC, and potentially higher latency than current hard drives.
Seagate has been pursuing dual-actuator drives because density increases have been causing a decline in performance per TB, making the performance characteristics of big hard drives relatively more tape-like. It's a first small step in the direction of your proposals, but it doesn't seem to be a big hit so far. That indicates taking those ideas to their logical extremes is probably very uneconomical.
Thanks for your analysis, I appreciate it. I was kinda saying, "hey why don't we have LaserDisc sized hard drive platters, they would be HUGE!". Using lots of smaller disks has so many advantages after I thought about it. One being additive bandwidth, if a single disk is 0.2GB/s if I need 400GB/s I know I need at least 200 disks to hit average write rates of 0.4TB/s.
Speaking of tape like behavior, it would seem that SMR would benefit a whole bunch from multiple actuators. It seems like rather than having dual actuators, as that doubles the number of simultaneous streams, it would help with lots of shared workload, but really that problem should be solved somewhere else -OR- widening the read head by some multiple. Instead of using a record player needle, use a rake. Interleave the tongues so there are like k tracks between splines. Also reduces the actuator travel distance.
We should think of individual drives not as DIMMs but the individual chips themselves. They can be remixed into what ever your workload demands.
HDDs had larger platters and multiple heads, they evolutionarily lost them. I could be completely wrong but judging by audible noises disks make I suppose vibrations make it hard to actually run multiple heads.
For the latter, it should suffice to use a monolithic closed loop control that intrinsically compensates for induced torque coupled between actuators through the base.
the $/Tb is far less with HDDs than SSDs and that doesn't show any sign of changing soon. HDD can go as low as 0.014$ per gb, while SSDs at their lowest are 0.070$ - 4x more storage for your $ and it also doesn't expire due to writes.
*edit? I am replying to myself in hopes of replying to everyone who responded.
You are all right and much smarter than me. I kinda outlined that lots of aggregated 3.5" (9cm) drives are better than some sorta boomer 30cm platters. I'll admit it was kinda for aesthetics.
But all the buzz seems to be about RISC-V and Apple’s chips, and Intel’s woes. All the hardware superstars are CPU people. Who’s the Jim Keller of storage?