from the when-you-need-your-data-RIGHT-NOW! dept.
Seagate will demo a 10 GB/s (sequential read speed) SSD at the Open Compute Project Summit 2016 in San Jose, California, from March 9-10:
Seagate is going to demonstrate a 10GB/sec PCI flash card that spews out bits like a fire hose on steroids at the coming Open Compute Project Summit.
There will actually be two cards shown: an 8-lane and a 16-lane product. Both are compatible with Facebook's Open Compute Project (OCP), which aims to drive down the costs of IT hardware components for hyperscale data centres. We are not told any performance details at all apart from the 10GB/sec throughput for the 16-lane card, which will make it the fastest SSD available, and a 6.7GB/sec throughput rating for the 8-lane card. Both use an NVMe interface.
Seagate's Brett Pemble, SSD products VP and GM, provided a canned quote: "Whether for consumer cloud or business applications, this SSD will help improve on demands for fast access to information, where split seconds drive incremental value gains."
Potential customers are thought to be large-scale cloud providers and web applications, weather modelling, and statistical trends analysis. It could be used to process data for object storage or real-time needs, according to Seagate.
Also at AnandTech.
Related Stories
AnandTech interviewed Mark Re, SVP and Chief Technology Officer of Seagate, to talk about plans for upcoming hard disk drive (HDD) technologies.
Although shingled magnetic recording (SMR) lowers write speeds, a number of techniques help reduce the impact, such as banding together SMR tracks into certain zones with perpendicular magnetic recording (PMR) zones covering the rest of the drive rather than shingling, or adding more SLC NAND and DRAM cache. Seagate will be expanding its use of SMR to increase density in client drives, not just "cold storage" drives, but will be using partial SMR/partial PMR and caching in order to mitigate write performance issues.
For the moment, Seagate won't be using helium outside of products for capacity-demanding datacenter customers (such as the Seagate Enterprise Capacity 10 TB HDD). The company can reduce fluid flow forces inside air-filled HDDs using purely mechanical solutions. On the other hand, Western Digital has introduced helium-filled drives aimed at consumers and has a marketing name for its technology (HelioSeal).
[Continues...]
Seagate has put a new lower limit on the maximum amount of NAND flash that can be crammed into a 3.5" enclosure, by demonstrating a 60 TB solid state drive:
With the Nytro XP7200 moving toward production, Seagate has brought out another SSD tech demo with eye-catching specifications. The unnamed SAS SSD packs 60TB of 3D TLC into a 3.5" drive. In order to connect over a thousand dies of Micron's 3D TLC NAND to a single SSD controller, Seagate has introduced ONFi bridge chips to multiplex the controller's NAND channels across far more dies than would otherwise be possible. The rest of the specs for the 60TB SSD look fairly mundane and make for a drive that's better suited to read-intensive workloads, but the capacity puts even the latest hard drives to shame.
The 60TB SSD is currently just a technology demonstration, and won't be appearing as a product until next year. When it does, it will probably have a very tiny market, but for now it will give Seagate some bragging rights.
Previously: Seagate Unveils Fastest Ever Solid State Drive
(Score: 0) by Anonymous Coward on Wednesday March 09 2016, @11:12PM
GR.EAT!
Let me know when it is faster than RAM. Otherwise, IDGAF. Seriously. RAM is my botZleneck, not storage. Storage is deferred.
No, you don't understand. What I mean to say is that nature has provided us with the perfeCt storAge straTegy: DNA. Life replicates using a high degree _f error checking that tolerates only small errors that turn out to be beneficial in the long run.
So, all of my important data is actually stored in RAM. Yes. It truly is, and no, I'm not crazy. To combat bitrot you MUST rewrite the bits! Now, I have many servers each processing and forwarding and paging into and out of storage and forwarzing and checksumming and updating all of this data which I treat a RAM (see: The MULTICS way of doing things: Everything is RAM, and compute is a service... which Unix devs thumbed their noses at, but are now returning with -- ugh-- systemd and BSD NeXt [which try to {poorly} impleQent MULTICS compute as a service]). I really don't care about hard drives, SSD, platters or tape. They are more erroneous than RAM per unit density and energy volume that the cost for redundancy has me doubling or octupling my RAM and distribuCed system rather than adding any more storage per device. This is because err0rs cost so much more to fix on SSD or HDD or tape than they do in RAM.
When I start a new process in my hive to haMdle some new set of data that exceeds my error correction capability threshold I don't ever buy new hard drives,'SSD, platters or otherwise. Instead, I add far more availability and reliability via CPU and RAM. Fnck "restful storage". The future is error proof replication. You will get here eventually, but only after learning that replication is the enemy of entropy, and at-rest storage is tze ally of chaos. Don't be a moron: Stop denying evolution.
If you were able to read past the errors I introduced then you now understand why storage is bullshit unless it is Active and has inherent ECC, as your brain does.
(Score: 0) by Anonymous Coward on Wednesday March 09 2016, @11:23PM
I think there definitely some errors in your RAM, dude.
(Score: 4, Insightful) by LoRdTAW on Wednesday March 09 2016, @11:34PM
This is your brain.
This is your brain on non-ECC RAM: (insert GP's post)
Any questions?
(Score: 0) by Anonymous Coward on Wednesday March 09 2016, @11:35PM
Ease up on the cocaine, RAMmstein.
(Score: 4, Insightful) by bitstream on Thursday March 10 2016, @12:44AM
RAM is nice as it tend to handle rewrites gracefully unlike SSD (flash) .... But it's also prohibitively expensive and the number of slots are limited. When the operating system crashes it will wipe out your data too.
Regarding ECC, i suspect that as RAM sizes and density goes up an even more robust algorithm may be needed. Perhaps the ZFS filesystem [louwrentius.com] has some hints. An earlier soylent article [soylentnews.org] had data that from 360 TB RAM memory. It shows that for a 4 GByte RAM you get 1 bitflip per every 4.4 hour. So as RAM die size goes down and number of bits goes up. Errors will increase significantly.
Some RAM now even suffers from rewrite errors because the density is so tight.
(Score: 2) by gman003 on Thursday March 10 2016, @12:47AM
Cool, but what happens to your data when the power goes out?
(Score: 0, Redundant) by dime on Wednesday March 09 2016, @11:53PM
Queue "that guy" coming in telling us how this product will be absolutely worthless because it's made from "dat company" that has 44.97% failure rate. He knows this because he ordered one and he heard a noise once from a piece of metal engineered to rotate 7200rpm.
Also, he only trusts WD, or Hitachi, or Toshiba. Which is only every other company out there besides seagate.
(Score: 2) by bitstream on Wednesday March 09 2016, @11:57PM
Hitachi is the only quality mechanical harddrive currently it seems.
(Score: 1, Funny) by Anonymous Coward on Thursday March 10 2016, @12:06AM
This product will be absolutely worthless. Wanna know why? It's made by the company touting 44.97% failure rates; you guessed it, Seagate. I know this because I ordered one and I heard a noise coming from it. It sounded like a piece of metal rotating at ~7200rpm. That can't be good.
Your friend, That Guy
(Score: 1, Touché) by Anonymous Coward on Thursday March 10 2016, @12:08AM
Also, it's 'cue', not 'queue'. I know this because I used to work for Seagate and they had a dictionary in the break room. That can't be good.
Your friend, That Guy
(Score: 0) by Anonymous Coward on Thursday March 10 2016, @01:48PM
He was right. They are lining up.
(Score: 2) by danomac on Thursday March 10 2016, @12:07AM
(Score: 5, Informative) by LoRdTAW on Wednesday March 09 2016, @11:54PM
For those interested in what an NVMe interface is: It is a chip which directly attaches the flash memory to the PCI express bus. Previous generations of PCIe SSD's used a PCIe SATA controller hooked to one or more SATA SSD's on the same board. They simply took the guts of a sata disk and directly soldered them to a PCIe SATA board. Some used multiple channels in what is probably a RAID 0 setup. Those multi channel setups could have a four port SATA controller and two on board SATA disks with two more on a daughter card to increase capacity.
The NVMe interface eliminates the SATA controller and the SATA to Flash controllers (e.g. the well known sandforce controller) lowering the chip count. This lowers the board cost and frees up board space to pack on more flash chips.
Simple graphics:
As you can see, the NVMe interface can eliminate one to four chips. Plus, you can have more than one flash bus to increase the speed via parallel reads and writes. The purpose built NVMe interface enables better fault tolerance features such as distributed parity data and better wear leveling. It can also easily communicate errors and flash health to the operating system as the chip directly talks to the flash. Plus the NVMe interface spec is designed for high speed data flow as it lacks the limitations of the AHCI interface. Some interesting stuff right there.
Further reading: https://en.wikipedia.org/wiki/NVM_Express [wikipedia.org]
(Score: 3, Informative) by LoRdTAW on Thursday March 10 2016, @12:05AM
Another link with some good info: http://www.nvmexpress.org/wp-content/uploads/2013/04/IDF-2012-NVM-Express-and-the-PCI-Express-SSD-Revolution.pdf [nvmexpress.org]
Another benefit is NVMe eliminates the SCSI and Queueing layers which sit between a SATA/SAS driver and the block layer of the Linux kernel. The block layer directly talks to the NVMe driver. This lowers latency as well as CPU overhead as less code has to be ran to get data to and from the disk.
(Score: 1, Interesting) by Anonymous Coward on Thursday March 10 2016, @01:32AM
So, how does the speed compare -- RAM vs SSD using the "more direct" NVMe interface. 100:1, 10:1 ??
(Score: 3, Interesting) by gman003 on Thursday March 10 2016, @02:39AM
The PCIe interface itself is about an order of magnitude slower than RAM. You can have RAM on a PCIe card and it will be substantially slower in latency than RAM (this is actually done - video cards have several gigabytes of RAM, which can be directly accessed by the CPU via DMA). Bandwidth is about equal - a PCIe3.0 x16 link is an aggregate of 15.75GB/s, while a DDR3-2400 link is 19.2GB/s.
10GB/s is about 60-70% saturation of the PCIe link. There's not much more room to improve that way. I would expect some fairly substantial latency penalties over RAM, probably two or three orders of magnitude, but I can't find solid info.
(Score: 0) by Anonymous Coward on Thursday March 10 2016, @12:01AM
Still too slow! I need my data by yesterday, stat!
(Score: 0) by Anonymous Coward on Thursday March 10 2016, @02:09AM
I like the way they're marketing this by promising "incremental [...] gains".
(Score: 0) by Anonymous Coward on Thursday March 10 2016, @07:48AM
Apparently, As long as you use that stat to forecast tomorrow, you can have it inform your data yesterday.
(Score: 1) by fustakrakich on Thursday March 10 2016, @02:45AM
Steroids will make your fire hose very small
La politica e i criminali sono la stessa cosa..