Seagate Debuts World's Fastest NVMe SSD With 10GBps Throughput (hothardware.com)

MojoKid writes: Seagate has just unveiled what it is calling "the world's fastest SSD," and the performance differential between it and the next closest competitive offering is significant, if their claims are true. The SSD, which Seagate today announced is in "production-ready" form employs the NVMe protocol to help it achieve breakneck speeds. So just how fast is it? Seagate says that the new SSD is capable of 10GB/sec of throughput when used in 16-lane PCIe slots. Seagate notes that this is 4GB/sec faster than the next-fastest competing SSD solution. The company is also working on a second, lower-performing variant that works in 8-lane PCIe slots and has a throughput of 6.7GB/sec. Seagate sees the second model as a more cost-effect SSD for businesses that want a high performing SSD, but want to keep costs and power consumption under control. Seagate isn't ready yet to discuss pricing for its blazing fast SSDs, and oddly haven't disclosed a model name either, but it does say that general availability for its customers will open up during the summer.

Comments

Inconceivable

[Anonymous Coward]

Hard to grok. You can fill up that new Samsung 16TB drive in 2 min 40 sec.

Re:Inconceivable

[kamakazi]

Fill it from what? /dev/urandom isn't even that fast on any normal hardware, and it would take a lot of spindles and a dedicated 100Gb/s network card to fill that pipe. This thing isn't practical for anything in a normal datacenter. The only place I can see something like this currently being a justifiable purchase would be as caching drives in a massive data acquisition system, like the LHC or similar, or very large scale modeling, like weather. I am actually curious about the capacity of these drives, is it going to be cheaper than similar quantities of RAM? Some of the applications that make sense are also candidates for large RAM caches, which are just as fast.

It is interesting to consider a machine with no RAM, this drive is faster that the slower DDR3 rates, which at first glance would mean that someone could design a truly non volatile computer, all state information except for processor registers is non volatile, so enough capacitor on board would allow a machine to write current pipeline contents to the SSD on power down, and it could resume exactly where it was on power up, only costing a CPU cache flush.

Of course TFA seems to indicate that it is pretty power hungry, so that kinda dampens excitement about a instant on/off device since most of the uses I see for that include power saving, either for unreliable power (off grid stuff) or long battery life, like a laptop with instant zero power hibernation.

As cool as this is, it isn't going to make a difference for normal PCs for at least a couple product cycles. It is fun watching the ongoing convergence of RAM and disk storage speeds and prices. There is a future in view now when we won't need to differentiate.

Re:Inconceivable

[St.Creed]

Well, we're just upgrading our SAS VA installation to 2TB of ram, and upgrading our SQL Server production db to 768GB of ram - it's pretty cheap nowadays and certainly much cheaper than optimizing a lot of queries - and we have better things to do than optimize queries.

That 16 TB SSD would be pretty neat too - I can already see how that would benefit the logs and tempdb on our installation. And 16TB is a bit too large for just tempdb and logs, but the 10Gbps is cool and I would certainly like the 15TB when we need to partition things or migrate tables, or build indexes. Our SAN is usually a bit faster than that (with 2 fibrechannel connectors) but it's cheaper to just plug this in for non-critical stuff and our SAN tends to be rather heavily used so sometimes it's a bit slower - I would like something that has no other users so I can get reliable performance metrics apart from outside influence.

Re:Inconceivable

[Gondola]

With a ~1-second Sleep command that isn't buggy as shit, I'd actually turn my PC off every night.

Re:Inconceivable

[pak9rabid]

/dev/zero of course!

Re:Inconceivable

[KGIII]

I'm thinking modeling with numerous disparate inputs into a dedicated array with multiple I/O ports. It's still going to hit the bottleneck but you can probably push and pull pretty fast - multiples of these would have been a godsend back in the day where the disk was the bottleneck and not the bus. Now, the bus is in the way and that's improving, slowly. This sits right there almost on the board. You should be able to slam it with multiple I/O and be able to (reasonably) get a decent bi-directional data stream.

Stuff like this makes my testicles wiggle as I'd have loved shit like this in about, say, 1998. Well, it'd have been much to fast and I'm not sure it would have had anything to plug into but you get the idea.

Re:Can anyone explain that speed in football field

[Anonymous Coward]

Imagine a 10GB picture of a football field. This SSD can transfer ONE of those pictures per second!

Re:Can anyone explain that speed in football field

[avandesande]

If you break a cookie into 10 pieces that each represent a GB and drop them from 4.9 meters into a box in the middle of a football field, it is like that.

Re:No model name?

[Joe_Dragon]

This is for servers not desktops any ways the desktop systems are limited in pci-e lanes and a high end system will have 2 or more X16 video cards + an X4 ssd.

Decisions decisions

[tlhIngan]

Crap. Now what to do here for my new PC build?

Most motherboards with 2 or 3 x16 slots really only have all 16 lanes hooked into one slot - the others are usually 8 lanes or less - 16-8-4 iisn't even an uncommon configuration (PCIe tip - the slots are really just physical - you can put x16 slots even though it's hooked up to x1 so you can fit in any PCIe card, albeit only running at x1 speeds. It's why Apple's old Mac Pros used x16 slots - that way they can accept ANY PCIe card).

So now what to do... GPU in x16 slot, and slow down my fast SSD by putting it in a x8? Or have my SSD be nice and fast by putting it in the x16 slot and slow down my FPS by putting the GPU in the s8 slot?

Nevermind if you want to do SLI or CrossFire and now have to deal with 2 x16 GPUs and 1 x16 SSD...

Re:Decisions decisions

[Kokuyo]

Dude, any GPU will do just fine at 8x. Or how do you think SLI would work otherwise? The beefiest gamerboards have 20 PCIe lanes max.

Re:Decisions decisions

[Anonymous Coward]

Try this board from 6 years ago:
http://www.overclockersclub.com/reviews/asus_sabertooth_x58/3.htm

Now consider that there are systems with 2 X58's in them.

That said GPU's typically need fast interfaces between onboard memory and the GPU. Filling textures from an HD or system ram only has to take place once per level (or whatever), so speed there isn't very important.

Re:Decisions decisions

[threephaseboy]

Intel boards with LGA 2011-3 sockets have 40 PCIe lanes available coming off the CPU.

Re:Decisions decisions

[DRJlaw]

So now what to do... GPU in x16 slot, and slow down my fast SSD by putting it in a x8? Or have my SSD be nice and fast by putting it in the x16 slot and slow down my FPS by putting the GPU in the s8 slot?

Simply, no.

You are right about the physical PCIe slot connectors.

You could be right about the physical PCIe slot wiring, but in many boards with two x16 slots (assuming that there are only two, with the chipset-wired slots not being physically x16) both are electrically wired to be x16. You do not have to put a card in a specific one of the two slots to have an operational x16 connection.

You are wrong about the functional connections. I'll assume an Intel-compantible motherboard since that is what I'm familiar with. AMD-compatible motherboards could be different - I simply do not know.

Intel CPUs provide 16 PCIe lanes for connection to the x16 slot(s). If you have one card inserted, that slot will be allocated all 16 lanes. If you have two cards inserted in a board providing two slots, each slot will be allocated 8 lanes. In Z170 boards there could be three CPU-connected slots, and with three cards inserted in such a board, the slots would be allocated x8/x4/x4. See here.

Everything else runs off chipset-provided PCIe lanes, which are connected to the CPU by a PCIe x4-like . Thus, for example, in my Ivy Bridge system (Z68), there is a third PCIe x16 physical slot that is PCIe x4 electrically wired and functionally PCIe x1-connected unless I set a BIOS option that disables certain other peripherals (USB3 and eSATA add-ons).

If you connect your GPU and this SSD at the same time, you will be either x8/x8 (if using CPU-connected slots) or x16/x4 (if using one CPU-connected slot and one chipset-connected slot). That x4 would also be shared with every other I/O connection in the system due to the DMI "x4" like bandwidth limitation.

PCIe PLXs switches add lanes to slots, but do not add further connections to the CPU or chipset. At the end of the day you're sharing either the 16 CPU-provided lanes or the 4 chipset provided lanes in Intel's consumer-oriented boards. You have to go to the LGA2011 socket and workstation chipsets to gain more available bandwidth to the CPU.

Re:Decisions decisions

[silas_moeckel]

Get a socket 2011 mb, it's not like these are meant to go into a single socket board.

Re:Decisions decisions

[pjrc]

If you can afford this "enterprise" SSD, you can certainly afford a Xeon or Haswell-E and LGA2011 motherboard with 40 PCIe lanes.

Re:Decisions decisions

[Kjella]

If you can afford this "enterprise" SSD, you can certainly afford a Xeon or Haswell-E and LGA2011 motherboard with 40 PCIe lanes.

Yeah. The nice thing about x16 cards is that you can probably reuse graphics-designed systems. Like this one, four x16 slots in a 1U chassis, 2-way system so you have 80 lanes total:

http://www.supermicro.com/prod...

Drop in four of those cards and you'll have a pretty decent database server, I imagine...

Re:Decisions decisions

[DRJlaw]

What you state matched Intel chips years ago. Then in 2015, Haswell-E CPUs came out supporting 40 cpu lanes. Each x16 slot on those boards is allocated a full 16 lanes.

It helps if you read what was written before you reply.

"At the end of the day you're sharing either the 16 CPU-provided lanes or the 4 chipset provided lanes in Intel's consumer-oriented boards. You have to go to the LGA2011 socket and workstation chipsets to gain more available bandwidth to the CPU."

Is Haswell-E a consumer-oriented platform? No. Is it LGA2011 socket? Yes. Did I mention that it gives more available bandwidth? Yes.

Also the DMI bus on Skylake generation is now a x8 PCIe link not a x4.

No. Pre-Skylake was DMI 2.0 (PCIe 2.0 x4 equivalent) and Skylake is DMI 3.0 (PCIe 3.0 x4 equivalent). Any PCIe slots coming off the consumer chipsets are x4 or less.

I challenge you to find an electrical x8 slot coming off any Intel consumer chipset, 100 series or earlier. I would hedge and say without a PLX chip being involved, but I don't think any manufacturer is insane enough to hang a PLX chip off the chipset instead of the CPU-provided lanes simply to provide a wider I/O card slot.

Re:Decisions decisions

[DRJlaw]

No comeback for the socket question, eh?

The real downside?

[bobstreo]

Seagate has terrible MTBF rates.

http://arstechnica.com/informa...

Re:Can anyone explain that speed in football field

[darkain]

Why would you need two of them though when you got good ol friendly /dev/null to copy all your files into!

Re:Can anyone explain that speed in football field

[orledrat]

TFA clearly states that it is damn fast:

The company is also working on a second, lower-performing variant that works in 8-lane PCIe slots and has a throughput of 6.7GB/sec (which is still damn fast).

Crucially, does anyone know whether I can safely use this with my "System D" system? It molested my cat last month, twice, so I want to make sure all is safe before slotting this sucker in.

We advertising, bragging..but...

[evolutionary]

We have no model number, no pricing, and no precise release date. Sounds like these tests are preliminary. Sounds like "beta" to me. I've been the the victim of an entire batch of enterprise drives which had experimental firmware (which actually shut down the drives at specific date!) so I'd take this announcement as market priming and take it was a grain of salt.

Re:Seagate platters != SSDs?

[Anonymous Coward]

That's for spinning platters - perhaps in an effort to get users to switch to "more reliable (only 10% failure rate) SSD from Seagate"

Mind you, I haven't seen any actual failure rates for Seagate SSDs, I didn't even know they made any pure SSD-only drives. They're best known for that horrible hybrid contraption which can likely easily conflate high SSD and high mechanical platter failures!

Re:Like Obama Endorsing TRUMP 2016

[smooth wombat]

Why would I be proud of someone who thinks bankruptcy is the first option, that taxpayers should foot the bill for private companies, who ran their casinos into the ground while everyone else was flourishing, who has been married 3 times and whose grasp of reality seems to be on par with Jeb Bush's, "My brother protected this country"?

Re:Can anyone explain that speed in football field

[orledrat]

TFA clearly states that it is damn fast:

Do note that "damn fast" is NOT equal to "ramadan fast".

There's differences: it takes less long than ramadan fast, to start with, and the cache-miss penalty is less severe per byte.

Re:Can anyone explain that speed in football field

[NotDrWho]

Is that an Imperial or metric fortnight?

Re:No model name?

[evolutionary]

Uh, I did I mentioned my batch of seagate ENTERPRISE drives had experimental firmware that shut down the drives at a specific date? My point of mentioning that was Seagate doesn't even test it's Enterprise hardware well. (Experimental firmware on a whole BATCH?!?) That is why I stopped using them. Enterprise doesn't seem to matter when it comes to quality control with Seagate.

In any event, for enterprise hardware. Oh, FYI, most server motherboards in my experience have 2-3 X16 PCI-ex slots. Some only have 1. So you can only have RAID 1, or 6 (5 has issues, suggest not using Raid 5). RAID 10 have to use other conventional sata connectors. I've rarely seen a desktop motherboard with less than 2 x16 PCIex slots.

Re:Transfer rates not bottleneck

[Billly Gates]

You will see zero benefit when booting pc and running standard programs. What really drives speed is IOPS and cpu (after mechanical disks and value ssds are gone). Tomshardware did a benchmark with raid 0 ssds vs standard ssds back in 2013.

They booted slower than non raid. Game loading speed didn't make a different either. BUT winzip and transfering a large 2 gig files where crazy fast.

So a server would benefit maybe but I doubt most have 10 gbs ethernet to come close. So unless you work on databases (those are IOPS, not bandwidth) but even then an enterprise SSD in raid 5 with several disks would approach the speed for queries even if the bandwidth is not as great. These are niches.

So if I were you unless you make 6 figures I would not invest. I would get a nice Samsung pro 850 SSD or raid 0 them for 1/3 the price. You wouldn't notice the difference at all

Re:Can anyone explain that speed in football field

[ausekilis]

If one carrier pigeon can carry a 1 MB floppy across a football field in 10 seconds, the pigeon's bandwidth is approximately 100KB/s. You'd need 100 pigeons carrying floppy drives across the football field to equate to 1 spinning platter drive (10 MB/s).

For a typical SSD, that means 3,000 pigeons carrying floppies (and a lot of interns to manage the disks). (300MB/s)

For this SSD, you'd need 100,000 pigeons carrying floppies. Or you could just be smart and go with 10 pigeons carrying 10GB thumb drives, but that's not how things are done around here.

Re:Can anyone explain that speed in football field

[Holi]

Sure as long as you ignore the time it takes to copy the data on to and off of the 8GB sd card.

Re:Seagate? No thanks.

[undefinedreference]

This is Seagate you're talking about here... We're not dealing with a flash memory manufacturer like Samsung, Intel, Micron, SanDisk, or Toshiba, so who knows what grade F- stuff Seagate has purchased for the lowest price possible.

Re:No model name?

[Joe_Dragon]

http://www.supermicro.com/prod...

http://www.supermicro.com/prod...

IOPS or bust

[Effugas]

Talk to me about IOPS -- Input/Output operations Per Second -- or don't pretend you're talking to me about performance.

Not saying this isn't actually really exciting, but that's the metric in at least 90% of use cases.

End of RAM vs. file system?

[Mal-2]

When will it become practical to eliminate the difference between temporary storage and long-term storage and just "execute in place", using RAM as a disk cache? It sounds like the speed is there already. If the storage is dangling off the memory controller rather than the PCIe controller, that would eliminate the worry about "lanes" as well.

Re:End of RAM vs. file system?

[UnknowingFool]

Well given that Flash based SSDs have an limit on the number of writes, using Flash for memory operations is not practical in the long term.

Re:End of RAM vs. file system?

[Mal-2]

This is why I mentioned "execute-in-place" specifically. If RAM is provisioned for working data sets but programs can simply be dumped at any time (because the pointer goes to the storage, not to the RAM), then the flash would take less wear than currently. Rather than swapping things out, they just get flushed and re-read as necessary rather than the current paradigm of "load from disk, execute from RAM".