Slashdot Mirror
Samsung Starts Mass Producing an SSD With Monstrous 30.72TB Capacity (betanews.com)
Brian Fagioli, writing for BetaNews: Samsung says it is mass producing a solid state drive with monstrous capacity. The "PM1643," as it is called, offers an insane 30.72TB of storage space! This is achieved by using 32 x 1TB NAND flash. "Samsung reached the new capacity and performance enhancements through several technology progressions in the design of its controller, DRAM packaging and associated software. Included in these advancements is a highly efficient controller architecture that integrates nine controllers from the previous high-capacity SSD lineup into a single package, enabling a greater amount of space within the SSD to be used for storage. The PM1643 drive also applies Through Silicon Via (TSV) technology to interconnect 8Gb DDR4 chips, creating 10 4GB TSV DRAM packages, totaling 40GB of DRAM. This marks the first time that TSV-applied DRAM has been used in an SSD," says Samsung.
Then I looked at the fourth significant digit. it is 2. Yes, it is actually 30.72 TB. That 651 parts per million more than my what I originally thought. Now I am all ears, looking at it carefully, camping outside Alibaba container terminal to be the first one on the block to get it.
Very well done Dear Headline Writer, always provide very precise information. Next time, why stop with the fourth significant digit? You could be even more amazingly accurate and provide six, seven... why not go all the way to 11 significant digits! Most people have just 10 digits, so go for 11, that is a good number hard to beat.
sed -e 's/Chuck Norris/Rajnikant/g' joke > fact
I got a quote for these. Only $28k. Pretty good job by Samsung. The 16TB previous gen are about $11k.
then a new station wagon I could upgrade from all those damn tapes.
At 1700 MB/s the right question is: How many will it take to keep up with the current growth rate of the LOP.
John McAfee 'It was like that time I hired that Bangkok prostitute; to do my taxes, while I fucked my accountant'
RAM is far faster than flash, and far more tolerant to being written lots. Itâ(TM)s used as a cache to both speed up the drive, and reduce wear on the actual flash.
Is this a slow 5200 RPM drive, or does it actually spin at 7200 RPM or even enterprise-level 15k rpm?
Currently cheaper to get a three pack of spinning 12TB drives (totalling $1350).
No one is really buying SSD in the enterprise based on it's price vs HDD.
I browse on +1 so AC's need not respond, I won't see it.
A car and a leg!
Or two legs, even!
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
Since Samsung was the one who engineered stackable V-NAND, I suspected that they could multiply the current market's SSD capacities by the layer-count. Maybe that's not what's happening here, but it sure seems like Samsung will have the edge in increasing SSD capacities.
"PM1643, why aren't you at your post? PM1643, do you copy?"
William George
eventually with the deepfakes type stuff, we won't need LOP's anymore, just a program to generate on the fly whatever midget+monkey+donkey fetish the mind can conjure up.
Coupled with VR, this will spell the end of mankind -- constant, ever increasing levels of titillation and depravity.
SAS/SCSI needs to go away.
I'm sick of "enterprise" gear being a shitty HBA to a shared storage device that slows you down (both speed and latency from the HBA, and the latency and shared aspect part of the storage device).
The world has spoken - NVMe PCIe flash isn't just the way of the future, it's the way of today.
But for servers I want it in a hot swap drive bay, and I need it to be redundant. There are only a handful of RAID controllers out there that will support such a thing. (And as far as I know, you're limited to 8 lanes of PCIe 3.0 with the latest "tri-mode" shit.)
Can we please get more support for NVMe PCIe storage in RAID controllers and backplanes? I've been actively waiting for this shit for about 2 years, and it's still a rarity and a mess. SAS/SCSI is dead. Let's bury it already. Fucking Epyc-based systems have 128 lanes of PCIe 3.0! Why don't we fucking use them?
Give me a backplane with 8 hot swap bays for 2.5" drives and pipe that shit over a 32 lane PCIe link (yes, it goes to 32, not just 16) to a fast RAID SOC. Each bay can get 4 lanes. And if you really want to be smart, you'll fucking let people connect M.2 devices on the backplane, not just U.2 devices. Because there are like 7 total U.2 devices in the world and they're all the same fucking thing as the M.2 equivalent in a 2.5" case but with a much larger price. Fuck that. I'll run servers on 960 Evos (or 960 Pros if I'm feeling paranoid). A M.2 to U.2 adapter, a 2.5" enclosure, and some thermal pads are all you need to get an affordable, fast, and reliable U.2 drive.
You can afford to spend a lot of time structuring your databases ...
The extrapolated cost of this 30TB drive is less than the burdened monthly cost of a single engineer. Furthermore, if disk I/O is your bottleneck, then moving your DB from HDD to SDD is likely to give you far more speedup than fiddling with your schema.
In my experience running IT for a large department the cost saved in not having to replace drives, and worse, restore data is worth the cost of SSDs, probably just in labor costs but also the additional peace of mind.
Also:
"For most corporate workloads, the acquisition cost of flash storage is still significantly greater than for HDD storage. However, when operating costs are factored in, the TCO for SSDs may actually already be lower than for equivalent HDD arrays. Use of SSDs reduces data center costs for power, cooling, floor space, rack space, and maintenance. And as SSD purchase prices continue to fall, the TCO disparity can only grow greater over time."
https://www.zadarastorage.com/...
And if your company is small enough that a single 30 TB SSD is enough to meet your entire storage needs, then you probably aren't big enough to be so desperate for speed that you would buy a 30 TB SSD. I'm expecting most of these will be used for the most frequently accessed data for companies that are Google-scale, not companies whose total data capacity needs are rivaled only by my home RAID array. :-)
But "fiddling with your schema" so that tables that get hit frequently are indexed on SSDs or DRAM drives (and, if necessary, stored in their entirety on those drives) means you get most of the speed win without the expense of storing your *entire* data collection on an SSD. The more you can separate out the frequently accessed data from the rarely accessed data, the easier it is to keep your costs sane. And the bigger your total storage needs, the more that design decision matters.
Check out my sci-fi/humor trilogy at PatriotsBooks.
I'm not sure the HBA is really the problem. I mean, you want I/O offload to some other CPU for RAID and logical disk management and some kind of physical disk channel multiplexing.
There's no reason these things are inherently slow unless you're buying low end stuff. Obviously you'll get better raw I/O with direct NVMe off the PCIe bus to the CPU, but as soon as you get into RAID it gets more complicated. Either you use software RAID at the OS level or you have to add a RAID controller to your PCIe bus. The latter is actually worse than a HBA IMHO because the controller and the drives sharing the common PCIe bus slows the whole bus down for traffic that only the RAID controller needs to see.
With a HBA, you just write blocks to one device on your bus and you let the HBA write those to disk on a separate bus. There's no reason an HBA couldn't have its own PCIe bus for NVMe drives.
The reason I'm skeptical we'll see anything other than hot-swap NVMe on the main system bus is that storage seems to be lurching towards hyperconverged where you wouldn't have more than about 12 drives on a physical system anyway and it's all software RAID.
I think there is an evaporating market for device-dense storage devices.
At this point it's really demand that's keeping the cost up. Between NAND and RAM, the fabs are all running full tilt and the memory cartel of Samsung, Micron, and Hynix doesn't seem all that interested in opening up new capacity. It's still a ways to go before it can hope to match HDD in price per GB, but it's not purely a technical hurdle anymore. I think you will see HDD demand start dropping in the coming year, and eventually that will push HDD prices up. I suspect in another decade or so SSD and HDD prices will reach parity, or at least get close enough that it won't matter.
I browse on +1 so AC's need not respond, I won't see it.
And if your company is small enough that a single 30 TB SSD is enough to meet your entire storage needs, then you probably aren't big enough to be so desperate for speed that you would buy a 30 TB SSD. I'm expecting most of these will be used for the most frequently accessed data for companies that are Google-scale, not companies whose total data capacity needs are rivaled only by my home RAID array. :-)
I suppose it's what you do. To a video production company 30TB is nothing. For random document storage it's quite a bit. For a database/source code repository it's pretty huge. We have quite a bit of database data that we need to keep around for archive purposes that are almost never used again, but in the end the relative savings of HDD vs SSD in a relatively expensive storage unit that still needs backups etc. and the resources to partitioning and archive things it didn't add up so we're all SSD now. It's a bit like starting to go through your photo collection to see if you can delete 10MB jpgs when you can go down to the store, buy a 8TB archive disk and store 800000 more. If you're doing it simply to save space and not for house cleaning it's not worth it.
Live today, because you never know what tomorrow brings
I bought my first flash SSD in 1994 (I think, maybe 1995). It was for my Psion Series 3, had a capacity of 128KB, and cost £30. It was a single cell, so you could write to it but not erase without erasing the entire disk, so you needed to do a backup and restore to replace all of the data. At about that time, a 1GB hard disk cost around £100-150 (I don't remember exactly what, but I remember being amazed when the price dropped to under 10p/MB a little bit later). Let's say £120, for the sake of argument. That meant that the cost of a hard disk, per GB, was about 8,000 times less than the cost of flash.
It's been a couple of years since I last bought any disks (my last two laptops have had SSDs, but the newer one is now over four years old and has a 1TB SSD). It looks as if 500GB SSDs now cost about £120, which is about the same price as a 5TB hard disk. That's a difference of about a factor of 10 in price - a big jump from a factor of 8,000. I'd expect that gap to close pretty soon.
Perhaps more important is the bottom end. You can get a 120GB SSD for about £30, or a 500GB hard disk for the same price. Note that now we're down to a factor of four difference in price. For a lot of users (e.g. corporate desktops where important files are stored on a server), the difference in utility between 120GB and 500GB capacity isn't that great, but the difference in performance between SSD and hard disk is a lot more noticeable, so it makes sense to go with the SSD.
This is the real problem for hard disk: if enterprise, mobile, and low-end desktops move to SSDs, where do you get the volume to pay for R&D on the next generation?
I am TheRaven on Soylent News