MAMR Breakthrough for Next-Gen HDDs | Hacker News Reader

MAMR Breakthrough for Next-Gen HDDs

anandtech.com

94 pointsvanburen36 comments

Mechanical storage is only efficient for bulk storage because NAND chips are still being price-fixed just like LCD panels and RAM chips before that. It's no accident it's the same manufacturers involved either. And just like those other products eventually saw a large international investigation culminating in "punishments" for global price-fixing, so will NAND. And when that happens, the fact hard drives have material costs that radically outstrip basic NAND chips (which are about as dead-simple as it gets) will very rapidly result in NAND storage offering tens or hundreds of times as much capacity for the same price. It's absurd the mechanical charade is being allowed to be perpetuated as long as it has been.

You can't make devices that deal with motors and rare earth magnets and spinning platters coated with ruthenium and other rare materials at insanely exacting tolerances, encapsulated in hermetically sealed Helium bubbles for cheaper than you can photolithographically lay out a bunch of NAND gates in cheap bulk semiconductors. There really isn't any word for it other than absurd. And the fact that NAND chips are in literally EVERYTHING means they are commoditized. Which means economies of scale make them cost almost nothing to manufacture. And yet... it still costs you 4x or more to get an SSD rather than several pounds of spinning metal? Nah, that's not how things work without help.

Why stop with microwaves? Why not make platters out of pure gold and the read/write heads out of synthetic diamond? Maybe integrate a cryogenic cooling system and store the data in a Bose-Einstein condensate? At this point it seems people will believe even that is cheaper than some NAND chips run off a line like printouts.

I admit to not knowing anything about how either drive type is made, but if what you are saying is true, then I wonder why other big players aren't just stepping up and making their own SSDs?

If Google can make their own processing unit just for running ML, why aren't they also making their own SSDs to drastically decrease storage costs and improve storage performance?

> I wonder why other big players aren't just stepping up and making their own SSDs

There are only a handful of fabs making the required NAND chips. Spinning up a new fab takes years and hundreds of millions of dollars, not to mention some serious technological and manufacturing know how. So it's really not easy for someone to just up and enter the NAND market.

I don't exactly doubt that price fixing is happening, but my understanding of current high SSD and even RAM prices at the moment is that there is a serious demand that outstrips the current fabs abilities. Mobile devices are eating up a lot of the NAND output.

In addition, chip fabrication processes are notoriously fickle, especially at very high densities. I wouldn't be surprised if 10% of the NAND chips came out DOA from the production line.

In the case of GPUs they can turn a defective GPU into a lower tier GPU by disabling malfunctioning components, which means that it isn't a total loss. I doubt NAND chips can be salvaged in the same way. Since they are so simple, there's nothing to recover. It goes straight into the bin.

The difference is that NAND doesn't need to BE high density except in those mobile devices. For both consumer grade SSDs and enterprise SSDs, NAND already has such a huge storage density advantage that if they actually used up the space available inside a 2.5" or 3.5" case and weren't concentrating on transfer speed as much, yields wouldn't be much of a problem.

NAND does actually have a degree of flaws it can tolerate as they are made today in consumer SSDs. I am not certain, but SLC Enterprise SSDs made for database servers and the like might get the best yield chips I'd guess. On consumer grade devices, there is an amount of 'slack' space that the chips actually can accomodate that is used for relocating data from damaged areas, wear levelling, some bookkeeping, etc. So if you buy a 1TB SSD, there might be enough actual storage on the chips to hold 1.1TB if all of it was made available. I'd not be surprised if particularly bad runs come out and get binned as 512GB devices because large portions of the chips are unreliable.

This is not especially dissimilar from how CPUs were made and marketed for years - CPUs that wouldn't clock at 1.6GHz would get sold as 1.4GHz, for example

Kinda like how ladders are rated: they say 300Lbs .. it'll probably take more than double that - but if it breaks when you overload it, the manufacturer can point to the rating and say, "you exceeded its spec"

> You can't make devices that deal with motors and rare earth magnets and spinning platters coated with ruthenium and other rare materials at insanely exacting tolerances, encapsulated in hermetically sealed Helium bubbles for cheaper than you can photolithographically lay out a bunch of NAND gates in cheap bulk semiconductors.

Citation required??

Western Digital's gross margins are 40%, while Micron's gross margins are around 46%. This suggests that the margins of both NAND chips and Spinning Hard Drive rust are roughly equivalent at current market prices (at least, within a magnitude or so).

Large-scale manufacturing turns out to be efficient. At the moment, prices suggest that Hard Drives are indeed far more efficient to make.

If your hypothetical were true, then we'd expect Micron's gross margins to be larger. From my understanding, the high-tech NAND gates have a relatively low yield. Yields on bulk semiconductors are no where near as good as the yields of Hard Drives.

I was a part of research on HAMR a few years ago. I remember one problem they were having was the heating lasers would burn out after a few writes. There was also research into spin torque MRAM, so it's interesting that spin-torque was used to make something that replaced HAMR.

Are data centers still using SATA to connect this big of drives? Seems like when your disks gets much bigger than 10GB, using SATA gets pretty limiting

why? do you expect mechanical drives to deliver >600MB/s peak transfers?

Even the 12TB drives are only advertising 250MB/s. https://www.anandtech.com/show/11903/seagate-ships-consumerf... That's a lot higher than the 150 MB/s or so you'd get from a 3TB drive, but it's not pushing the limits of SAS or SATA. Unless you cheaped out and connected a bunch of drives to a single controller.

Yes, and funny enough the reason we got 6Gbps SATA was because of SSD.

SAS 4 is the works and should be finalized soon. It will supposedly have 22.5 Gbit/s (2.8 Gbyte/s) of bandwidth. Whether or not a HDD can actually saturate this is suspect.

This is cool. But I assume SSD is going to get better too.

If SSDs were forecast to overtake magnetic drives, then WD would be unwise to invest here. But the graph from the article shows a 10x cost/GB continuing into 2028.

So, yes, SSDs will improve. But they exist in a different tier/market segment until they can close that cost gap.

Not just cost but density. I'm pulling these numbers out of the air, but suppose you could have a $100 4TB SSD or a $600 24TB HDD some years from now. Many datacenters will still prefer the HDD just for the rack space considerations.

I'd expect SSDs to be significantly more expensive but higher density actually.

My perception of the future is maybe a 20TB SSD for $2000 vs 15TB HDDs for $500.

SSDs are already more dense than HDDs and will likely remain that way. The only remaining barrier is $/GB.

I was all set to doubt you, but then I found this: https://www.micron.com/products/solid-state-storage/product-.... My ignorance is showing again.

There's even 15 TB available for sale today in the 2.5" form factor: https://www.cdw.com/shop/products/Samsung-PM1633a-MZILS15THM...

And 60 TB in 3.5": https://arstechnica.com/gadgets/2016/08/seagate-unveils-60tb...

No spinning rust drive comes close.

Of the currently shipping products SSD has the highest density. Samsung has a 2.5" 15 TB SSD. The closest magnetic storage is 12TB 3.5".

Another important factor to remember is that your HDD will give you 200 MB/s of sequential data. If you have a random pattern it's going to drop to 10 MB/s or less. (15K RPM SAS drives can do around 150-170 random IOPS, assume 32 KB per operation and you're looking at 6 MB/s.) SSD will push out 10x+ that rate for random IOPS. A 100 TB HDD sitting behind 200 MB/s is painful but sitting behind 5 MB/s is insane. There's a place for such devices, but it's not serving your OLTP databases or kafka message queues.

Exactly. HDDs really can't be compared to SSDs directly without considering the goals of the system. They are both storage devices, but the pros/cons make them like comparing a car to a truck. They both transport people and things, but the car gets better gas mileage than the truck but has a lot less towing capacity. The job you need done is a huge factor in which solution is the most optimal.

That 'cost gap' is completely artificial and supported solely by price-fixing. Remember when LCD panels only dropped in price a few cents a year despite being in tons and tons of products? That was price fixing. The same thing had occurred with RAM in the years before LCD. And it was the same manufacturers involved. This time around, the story is the same. A commodity product that costs almost nothing to produce yet somehow magically never drops in price by anything but the most marginal of amounts. When those manufacturers got busted for price-fixing it only took a couple years before the price of an LCD panel dropped to a quarter of what it was previously, and SSDs will drop even faster.

WD might expect the price-fixing will continue for several more years, though. The corrupt South Korean administration dropped their investigation into price-fixing in the NAND market several months back. I don't know what impact that will have on the general international effort to prove the collusion that seems so obvious.

I just want a 20TB SSD for $500. It's 2017 and no, that's not too much to ask. They're friggin parallel grids of NAND gates with a bunch of multiplexor/demultiplexors bolted on. All photolithography. I don't need them to be those tiny little M.2 sticks, go ahead and give me one in a friggin 5.25" case. At this point they've gone about a mile past where I got fed up with this crap.

I'm not sure what you are on about, the prices of SSD dropped by a factor of 3 or so from 2013 to 2016. That's more than HDD. At the same time, SSDs got faster and faster, while HDD speed is essentially constant.

> If SSDs were forecast to overtake magnetic drives, then WD would be unwise to invest here.

Why not both?

WD bought SansDisk for a reason. WD is likely going to be investing in both Hard Drives AND NAND Flash in the near future.

WD has limited R&D money. If your investment isn't going to pay out because SSDs surpass HDDs in GB/$, then you've just wasted a big up front investment (and opportunity you could have been investing in for SSD). The grandparent comment is hypothesizing that SSDs may surpass HDDs on GB/$ before the HDD R&D pays off.

Datacenters still are relying on spinning storage and pretty much everywhere I read will continue to for the foreseeable future.

I'm curious what the other factors are beyond just the cost per TB/PB in each rack. Like reliability over very long time scales... beyond the cost factor and into archival territory. Or in an emergency or destructive scenario (i.e. do platter drives fair better in a fire than SSDs)

The only reason they're using spinning storage is it's cheaper. I work for a big storage vendor. We would switch to all flash tomorrow if it were equal or lower $/GB (already has better performance and density).

no you wouldnt, Flash is useless for cold and very hot storage, anything that is not touched for more than 6 months, or will be rewritten more than ~2000 times in under a year is either impossible or uneconomical with solid state.

I'm not talking about the market segments where tape or RAM dominate (and would continue to dominate). I think the use cases you are imagining for spinning rust are a tiny portion of the market for storage, or simply do not exist. For example:

It's hard to actually rewrite spinning rust 2000 times a year. That's exceeding 100% utilization with completely optimal sequential write pattern on 7200 RPM drives. Say you can write 200 MB/s continuously[0]. That's about 6 PB/year. With a 6 TB harddrive, that's only 1000 full overwrites a year.

Maybe 10k or 15k RPM drives can overwrite 2000 times a year, but those have vastly higher cost than 7200 RPM drives. At some point, you just overprovision the flash enough to bring the number of required overwrites down, and it's still cheaper than 10-15k RPM drives.

[0]: Approximately average for enterprise drives: http://www.tomsitpro.com/articles/enterprise-hdd-testing,2-8...

Wasn't there some speculation that Glacier and co actually use optical media, not tapes, for their storage? Does anybody know?

I'd heard that speculation too, but do not have any actual knowledge of what Glacier uses. I don't deal with truly cold, offline storage at all.

So is this a reason to consider buying WD stock? This looks like they've completely surprised the market with this technology leap.

No, in my opinion. But not because I don't think it's a promising technology announcement or a good company, but because a casual reader of articles on HN is not positioned to compete with the professional analysts.

They mentioned that they'd be hosting an event with innovation announcements months ago. Some traders would have been buying WD stock in anticipation of that event, and wagering whether or not that included MAMR, HAMR, etc, and what each of those possibilities would mean to the company outlook.

The event and the article describing the MAMR breakthrough were posted yesterday at 9:30 AM PT. Your 'window of opportunity' to respond to the article was the few moments it took to figure out which of your guesses about what they'd announce were confirmed or refuted at the event. Even then, you may have been reading it literally by eye, while your competition was parsing it with NLP, using a gigabit connection and server located near the wdc.com servers, or perhaps using ones near the stock exchange.

http://investor.wdc.com/eventdetail.cfm?EventID=185452 https://wdc.app.box.com/s/2o32m71nthf84ozbz8wlwyo59uqxb6og/f...

As a result of this preparation, concurrent with the announcement and market open, the stock 'jumped' from 84.27 to 86.22, and a somewhat remarkable volume of 917k shares were traded instead of their daily average of 144k shares, but have since moved around a bit and are gradually settling in:

https://finance.google.com/finance?q=NASDAQ:WDC

In short: you're literally 24 hours behind the real market. Don't try to play it minute by minute based on Anandtech or HN publications.

But don't let that discourage you from buying companies based on what you believe their long-term outlook to be. Or maybe do that, and just buy into mutual or index funds instead.

I assume that the OP might have a more long-term strategy in mind. If they can bring this to market quickly, it certainly is an advantage over the other HDD manufacturers, which might hold for a while. That's not necessarily priced in yet.

Its highly unlikely that an algorithm, however sophisticated can project the outcome of such an announcement in such a short time.

But discussions like this on a news/analysis piece by a fairly technical new source combined with above average human intelligence will project a more likely outcome.

The stock may go up on this news like you mentioned. But it more often than not settles down. But if OPs time horizon is medium term(a few quarters) or more, then its worth investing on such news.

Reply on news.ycombinator.com