Intel's SSD 910: Not that impressive

[taltamir]

http://www.anandtech.com/show/5743/intels-ssd-910-400800gb-mlchet-pcie-shipping-in-1h-2012
The article salivates all over it but its just unimpressive.
This is NOT a single chip native solution; it doesn't have any features. It doesn't have built in RAID or even normal mundane single drive features like hardware encryption.
This is simply a 4 (6gbps) port SAS card without raid and 2 or 4 completely independent drives mounted on it on extra boards (probably standard SATA controllers since SATA drives can be connected to SAS ports).

Each of the 4 SAS drives is:
Size: 200GB (user visble, 224GB internal)
Random 4KB Read (Up to) 45K IOPS
Random 4KB Write (Up to) 19K IOPS
Sequential Read (Up to) 500 MB/s
Sequential Write (Up to) 375 MB/s

Two pack price is $1929/2 = 964.5$ per drive.
Four pack price is $3859/4 = 964.75$ per drive.

This just look unappealing compared to clamping together 4 other drives of superior performance and price with a 2/4 port non RAID board.

 

[beginner99]

So basically if I create a database on one of the visible drives I actually only get specs you mentioned for that drive and not the full 1000/2000 MB/s for sequential read.

Meaning like 4 crucial m4/vertex 3s/... on a SATA-3 extension card (assuming a usable SATA-3 controller) basically get the same performance for much, much less?

 

[KingFatty]

I thought it was weird that when you click on the article to read about the *INTEL* SSD, you are presented with a picture of the OCZ SSD? Is there a marketing relationship between Anandtech and OCZ?

 

[Morg.]

It is indeed a piece of crap, but it's the first Intel pciE SSD and as some only buy Intel, a very good product to sell.

On the other hand, I believe the non-raid0 approach can be interesting, enabling you to make more sensible raid configurations w/ the ssds - rather than the usual 4x raid0 / While you may not be able to replace drives on the SSD card, you may be able to keep your data if one of the SSDs fails - a very good approach to consumer-grade hardware for professional purposes.

 

[taltamir]

So basically if I create a database on one of the visible drives I actually only get specs you mentioned for that drive and not the full 1000/2000 MB/s for sequential read.

Meaning like 4 crucial m4/vertex 3s/... on a SATA-3 extension card (assuming a usable SATA-3 controller) basically get the same performance for much, much less?

yes, exactly.
The 1000/2000 is intel saying "well, there are 2/4 completely independent drives sold in a package so lets just add up their speeds together".
If you created a seperate database on each of the 4 drives, then maxed out each one, then the concurrent speed of the entire system would be the figures intel gave.

On the other hand, I believe the non-raid0 approach can be interesting, enabling you to make more sensible raid configurations w/ the ssds - rather than the usual 4x raid0 / While you may not be able to replace drives on the SSD card, you may be able to keep your data if one of the SSDs fails - a very good approach to consumer-grade hardware for professional purposes.

If they used a typical intel controller then it would have least had the option of RAID0 (and starting with v11.5 driver RAID0 with TRIM). It would have been really neat if you had the choice of JBOD, RAID0, RAID10, 2xRAID1, or RAID5... but you are not given any choice here its just JBOD and nothing else, which is kinda worse then the forced RAID0. I was hoping intel would make a special RAID0 hardware controller that passes on TRIM commands without needing a specialized driver, that would have been epic win and the holy grail of SSD tech.

And while I am all for software implementations but the issues with this drive are:
1. It is ridiculously expensive compared to the other things on the market.
2. RAID0 is not even an option
3. It is not in any way technologically impressive (they just threw existing commercial components from companies like LSI on a board)
4. Its performance per card is relatively poor.

 

[Morg.]

As i said , the drive sucks ^^

But still for a home-made box, considering how people like using Intel stuff only, running my own software RAID10 across these 4 Intel drives through a PCIe interface is an interesting option.

Also wtf. trim command in 2012 ? do we really need that when garbage collectors do the work ?

 

[philipma1957]

so 800gb = $3859.00


I have a mac mini server with a pegasus r6 cost was 1350 for the pegasus r6 with 6 1tb hiatchi hdds. (before the flood)

I purchased 2 samsung 512gb ssds for 699 each.

so 1350 plus 699 plus 699 is 2748. the mac mini server cost me 955 at amazon it has 2 500gb internal drives so I am at 3703 I paid 40 for 8gb ram and 50 for a t-bolt cable grand total of 3793



I have a raid0 with the ssds a 1tb osx drive.
I have a raid1 with 2 1tb hdds
I have a raid0 with 2 1tb hdds

I have a spare pair of 1tb hdds.
plus 2 internal hdds of 500gb
a 3 year warranty on all of it. all of it is bootable.

 

[taltamir]

Also wtf. trim command in 2012 ? do we really need that when garbage collectors do the work ?

TRIM is absolutely essential for GC to work properly.
GC can not and never ever ever will be able to work properly without TRIM.
This is why TRIM was ratified and adopted so fast. (I can't think of another communication standard that came about and was widely adopted faster).

Highly aggressive GC algorithms operating without TRIM butcher your drive's longevity with significant write amplification and require idle time to maintain performance.

EDIT: A correction, a potential exception for the above is GC that has FS specific algorithms that read the FS' MFT and based on it perform TRIM. However this carries its own bag of worms.
Enough that it is arguable about the "work properly" part.

Also, the properly part does not refer to "being fast".

 

[razel]

I don't understand what the fuss is all about. It's priced beyond what we would pay because this drive is not meant for the general pubic. Look at their target:

If you are actually trying to run a server off this because your db or server is I/O bound, then server admin pros have many other tools to use rather than just buying hardware. That's why really good DBAs make big bucks. They don't just put faster hardware onto a machine. Production machines are not their toys to OC or upgrade at will.

 

[taltamir]

If you are actually trying to run a server off this because your db or server is I/O bound, then server admin pros have many other tools to use rather than just buying hardware. That's why really good DBAs make big bucks. They don't just put faster hardware onto a machine. Production machines are not their toys to OC or upgrade at will.

So what? this does not justify inferior performance for more money.

How is it better? why would you buy it?

Also, google went and showed that large arrays of consumer grade hardware make more sense then overpriced enterprise hardware.

 

[ArchAngel777]

TRIM is absolutely essential for GC to work properly.
GC can not and never ever ever will be able to work properly without TRIM.
This is why TRIM was ratified and adopted so fast. (I can't think of another communication standard that came about and was widely adopted faster).

Highly aggressive GC algorithms operating without TRIM butcher your drive's longevity with significant write amplification and require idle time to maintain performance.

My 4 Intel X25-M G1's have used 1% of their write cycles. I have owned them for over 2 years. I am not worried that I don't have TRIM.

 

[taltamir]

My 4 Intel X25-M G1's have used 1% of their write cycles. I have owned them for over 2 years. I am not worried that I don't have TRIM.

The G1 (and for that matter, intel drives in general) do not have aggressive GC, it conserves the drive's lifespan at the cost of performance.
Aggressive GC sacrifice lifespan in order to try to maintain new state performance, which they can't do as well as TRIM but they can get pretty close if aggressive enough.

Also, I am editing in a correction

 

[Coup27]

My 4 Intel X25-M G1's have used 1% of their write cycles. I have owned them for over 2 years. I am not worried that I don't have TRIM.

The fact that your usage pattern has caused so little wear on 50nm NAND which isn't even available in consumer drives anymore is not really an argument against a technology as imporant as TRIM, especially if you reference it to a drive with 2xnm NAND with a much greater workload.

 

[razel]

So what? this does not justify inferior performance for more money.

How is it better? why would you buy it?

Also, google went and showed that large arrays of consumer grade hardware make more sense then overpriced enterprise hardware.

Never said it was better. Never said I would buy it. Never said there was any justification for inferior performance for the money.

Enterprises do not care for any of that. So why should you?

It's sad how reality TV confessional boxes lead to the 'me' generation which lead to narcissistic reality TV stars which lead to twitter and now alot of people find it easy to just spout opinion and wanting everything for less, instead of being productive about their ideas. Looks like the many of those in fatwallet are here who want unicorns and rainbows for free.

I want sunshine and easy days. It's beautiful outside, why bother about something that wasn't meant to be sold to you? If you want faster performance with NAND be a man, stop complaining and use that energy to contribute your knowledge to research.

 

[Concillian]

They are significantly better in a way that their target customers will find important: Write cycles.

It's ~an order of magnitude better in terms of write cycles compared to anything other than SLC NAND (which will not be cost competitive). It's multiple orders of magnitude better than any RAID + consumer SSD offering, which is typicallly in the <100TB range on writes.

This is a product for someone with write heavy workloads who would benefit from an SSD, but have avoided SSDs due to their workload would chew through a typical SSD's write cycle in a matter of days or weeks.

You don't have a workload like that, you probably don't have interest in that product. But there is definitely a market for an SSD with a ~10 petabyte total write capacity. vs <100 Terabyte capacity for a consumer level drive. Keep in mind when comparing consumer level drives that they'll need to be replaced about 100 times more often than a 910. 100 times more write cycles is a LOT.

 

[beginner99]

You don't have a workload like that, you probably don't have interest in that product. But there is definitely a market for an SSD with a ~10 petabyte total write capacity. vs <100 Terabyte capacity for a consumer level drive. Keep in mind when comparing consumer level drives that they'll need to be replaced about 100 times more often than a 910. 100 times more write cycles is a LOT.

Ok true. it has more durable nand than a standard consumer drive and the slide does specifically mention caching for a SAN where durability could get a big plus.

But I guess we are all just disappointed at how little innovation (= 0) this product has.

 

[razel]

Hey taltamir. My intent wasn't to target you or anyone else. You're nothing like what I described.

You've been EXTREMELY productive in these forums and have enjoyed reading your other posts. Sorry.

 

[taltamir]

Ok true. it has more durable nand than a standard consumer drive and the slide does specifically mention caching for a SAN where durability could get a big plus.

But I guess we are all just disappointed at how little innovation (= 0) this product has.

Pretty much.
This product is inferior in every way EXCEPT write durability. And from what I know of eMLC, the use of eMLC is not the CAUSE of this product being inferior in every way. It is also lazily designed (minimizing R&D budget)

I looked around and I don't see eMLC product except from intel, which has this offering:
http://www.newegg.com/Product/Produc...82E16820167079

Which is cheaper but slower. yet comes with built in AES 128-bit hardware encryption.

The product disappoint me because intel seems to have taken the approach of "no competition, lets be lackluster in every possible way" and it will cost them if we see other companies enter the eMLC market.

 

[Lonbjerg]

it doesn't have any features.

It has Intel's low, low RMA rate...way lower than anyone else....important part in storage

 

[Morg.]

They are significantly better in a way that their target customers will find important: Write cycles.

It's ~an order of magnitude better in terms of write cycles compared to anything other than SLC NAND (which will not be cost competitive). It's multiple orders of magnitude better than any RAID + consumer SSD offering, which is typicallly in the <100TB range on writes.

This is a product for someone with write heavy workloads who would benefit from an SSD, but have avoided SSDs due to their workload would chew through a typical SSD's write cycle in a matter of days or weeks.

You don't have a workload like that, you probably don't have interest in that product. But there is definitely a market for an SSD with a ~10 petabyte total write capacity. vs <100 Terabyte capacity for a consumer level drive. Keep in mind when comparing consumer level drives that they'll need to be replaced about 100 times more often than a 910. 100 times more write cycles is a LOT.

No sir, that is just wrong . consumer SLC drive raid >>>> this piece of crap.

And yes SLC is competitive considering the awful price tag Intel demands.

100 times more write cycles IS NOT A LOT when you have 8.000 write cycles and a wear-leveling algorithm that guarantees it across the whole disk.

Get over that write cycle myth already, it's been a while since that didn't matter, since write amplification was fixed close (or below) to 1.0 and since wear-leveling enabled drives to reach 90+% of their write cycles in a uniform manner.

 

[Morg.]

TRIM is absolutely essential for GC to work properly.

No .
TRIM is an instruction sent by the os to the drive controller for OS-managed GC.
Drive-managed GC doesn't require TRIM or support for TRIM even if it is quite clear that the controller must have understanding of the filesystem in order to decide which blocks are to be cleaned.

 

[Cerb]

TRIM is absolutely essential for GC to work properly.
GC can not and never ever ever will be able to work properly without TRIM.

So, without TRIM, if you write to logical block 200 5000 times, you will wear that physical block out? No. Wear leveling cannot function without garbage collection, and wear leveling does function without TRIM. SSDs work fine in Linux (TRIM is there, but I don't know any distro turning it on by default, yet), OS X, and in RAID. A wear-leveling algorithm implemented without garbage collection would stop writing as soon as the physical capacity of the NAND had been written to completely one time. It would be useless.

This is why TRIM was ratified and adopted so fast. (I can't think of another communication standard that came about and was widely adopted faster).

No, because it's just not needed for it. TRIM was probably adopted so fast because everyone introduced to it felt like doing a *facepalm* for not having thought to introduce many years earlier, because it's a plain good idea.

How is it better? why would you buy it?

Lack of faith in OCZ, and quite a bit of faith in Intel, for one thing, if I was managing rooms full of racks of computers limited mostly by their SSD RAIDs. Sure, there's no frills, but RAID 1, 0, and 10/0+1 can run fast enough in software, so that's not likely to be too much of an issue. Windows and *n*x software RAID is reliable, too, though there are caveats when it comes to using it for the OS volume.

For a home user, or workstation user, even, it's probably useless. Intel wants volume and margins, though. PCI-e SSDs just don't sell outside of servers. Sure, what are made sell, but it's like the big bad 300W dual-GPU video cards: tens of thousands. A small number of big customers could get Intel millions of units sold. You, with you audio, video, CAD, and games, are statistical noise.

 

[taltamir]

No .
TRIM is an instruction sent by the os to the drive controller for OS-managed GC.
Drive-managed GC doesn't require TRIM or support for TRIM even if it is quite clear that the controller must have understanding of the filesystem in order to decide which blocks are to be cleaned.

This is extremely wrong.
There is no such thing as OS managed GC.

TRIM lets the drive know what is junk so its GC can work properly. Read up on what TRIM actually does.

Drives aware of the FS are a whole different can of worms and not all GC works that way.

So, without TRIM, if you write to logical block 200 5000 times, you will wear that physical block out? No. Wear leveling cannot function without garbage collection, and wear leveling does function without TRIM.

Strawman. This is nothing at all like anything I said.

TRIM = Max performance
No TRIM with mild GC = slower performance
No TRIM with agressive GC = near TRIM performance at the cost of severe write amplification.

A wear-leveling algorithm implemented without garbage collection would stop writing as soon as the physical capacity of the NAND had been written to completely one time. It would be useless.

That is stupid. It wont stop writing it will just get very high write amplification.
And to prevent and mitigate this we have TRM as well as over provisioning.

Lack of faith in OCZ, and quite a bit of faith in Intel

Another strawman.
There are dozens of companies out there who make SSDs, why compare the one with the worst rep to intel?
Some have ironclad reputation on par with intel's.

For a home user, or workstation user, even, it's probably useless. Intel wants volume and margins, though. PCI-e SSDs just don't sell outside of servers. Sure, what are made sell, but it's like the big bad 300W dual-GPU video cards: tens of thousands. A small number of big customers could get Intel millions of units sold. You, with you audio, video, CAD, and games, are statistical noise.

And yet another strawman... this has nothing to do with any of my points or arguments.
I was only talking about its viability and usefulness from the perspective of a CORPORATION running a thousand units which NEEDS the speed SSD gives (and thus is willing to pay thousands of dollars more compared to HDDs) go for a product that is more expensive yet inferior in every way.
There has been one legitimate argument so far. It is cheaper then SLC, and has more write endurance the MLC, and there is just no competition in the eMLC market right now.

Also, I don't know where you people get this nonsense idea that corporations have infinite wealth and that they remain successful by burning it on useless stuff for the lulz.
Google is one of the greatest companies in the world and has a policy of using cheap consumer grade parts because the "enterprise" grade stuff is a ripoff.

 

[Cerb]

Strawman. This is nothing at all like anything I said.

TRIM = Max performance
No TRIM with mild GC = slower performance
No TRIM with agressive GC = near TRIM performance at the cost of severe write amplification.

No, what I quoted was exactly what you said. What is currently quoted above this text is not that. Also, the WA has not been severe.

That is stupid. It wont stop writing it will just get very high write amplification.
And to prevent and mitigate this we have TRM as well as over provisioning.

Do you consider 1.5-2 high WA? I don't. The differences just aren't that big*. Performance is aided much more than WA. That could change as we get to 1k, and even lower, write cycle ratings, but for now, the WA differences are minor.

Another strawman.
There are dozens of companies out there who make SSDs, why compare the one with the worst rep to intel?
Some have ironclad reputation on par with intel's.

Not a strawman at all. OCZ's are just under $5/GB, and widely available without being attached to a server, already, and without asking for a quote. OCZ does deserve credit for doing that, rather than relying mostly on middlemen and server vendor contracts. I hope Intel follows suit, and does the same, forcing everyone else to compete with each other out in the open, which a little company like OCZ simply does not have the power to do. Major memory companies do not exactly have a great history of wanting to compete on merit and value in open markets (and, I get the impression that LSI wishes it could be that way). Intel has a habit of being quite successful at it.

And yet another strawman... this has nothing to do with any of my points or arguments.
I was only talking about its viability and usefulness from the perspective of a CORPORATION running a thousand units which NEEDS the speed SSD gives (and thus is willing to pay thousands of dollars more compared to HDDs) go for a product that is more expensive yet inferior in every way.

More expensive and inferior to what other server PCI-e SSDs?

* Samsung's 470 being an outlier of the last generation

 

[taltamir]

No, what I quoted was exactly what you said. What is currently quoted above this text is not that. Also, the WA has not been severe.

You quoted what I said, then completely reinvented it to mean different things. Hence the strawman.

Do you consider 1.5-2 high WA? I don't.

This is not the WA you get with agressive GC.
This is the WA you get with mild GC that has poor "used state" performance on drives, the so called "degradation" SSDs experience.
I clearly stated it so stop making up arguments that I didn't make for you to shoot down.

Not a strawman at all. OCZ's are just under $5/GB, and widely available

OCZ being cheap or widely available doesn't make it not a strawman.
You elected to argue that intel is more reliable better then OCZ therefore this intel drive is not bad.
This is utter nonsense as there are dozens of manufactuers out there with better rep then OCZ.

More expensive and inferior to what other server PCI-e SSDs?

I clearly explained that could get SATA based on SSDs instead with no loss.
This so called "PCIe SSD" is a 4 port SAS non raid controller attached to 1 or two daughter boards, each daughter board is 2 SATA SSDs which plug into the SAS controller.

This is SLIGHTLY more compact than a SAS controller board + SATA SSDs. (Especially 1.8" ones)