Intel 910 Series 800GB PCI-E SSD Review

[SKYMTL]

Intel’s SSD division has been on a roll as of late. Their 330 series has offered entry level consumers the possibility of stepping up to an affordable alternative to spindle based media and for enthusiasts, the 520 series holds some of the best mass market SSDs that money can buy. Both have garnered a huge following and competitors have constantly found themselves playing catch up in the SO/HO and high end markets. However this refreshed lineup did leave a rather conspicuous hole since Intel’s Enterprise products lacked any meaningful steps forward. Well, that’s about to change with the introduction of the high performance, massively endowed 910-series.

In an effort to satisfy the needs of very demanding business IT administrators, Intel hasn’t pulled any punches here. By sporting a capacity of 800GB, a PCI-E 2.0 interface, four individual controllers and a single RAID chip the 910 800GB should pack enough performance to satisfy anyone. Indeed, its on-paper specifications look a lot like those of the OCZ RevoDrive 3 x2 but the two drives cater to completely separate markets. While the RevoDrive3 series is a prosumer model which is mainly meant for workstation environments and the 910 is a true enterprise grade storage device meant for servers and ultra high end workstations.

​

The difference in classification may sound subtle but unlike the OCZ option, the Intel 910 relies on extremely high end NAND modules and most importantly, is not bootable so you won’t be installing an OS onto it. Let’s not forget that being an enterprise-class product, it comes with a staggering pricetag: $1,929 for 400GB and the 800GB SKU tips the scales at nearly $4,000. This puts the 910 well outside the reach of most consumers but for the Enterprise environment, it isn’t overly expensive either.

While the price can be considered at best reasonable for its intended audience, the 910 does have many things which are clearly in its favor such as drastically reduced CPU overhead requirements and greater NAND longevity. On paper, this makes the Intel 910 a potentially great solution for the server environment and even workstations where the ability to boot an OS is of little concern. In order to prove its worth it will however have to deliver on its promises of stellar performance if there are any hopes of gaining traction with a consumer group that’s been notorious for their conservative approach towards new technology.

 

[SKYMTL]

A Closer Look at the Intel 910

Introducing the Intel 910

When engineering began on the 910-series, Intel’s main priority was to make an enterprise grade PCI-E storage device which was not only extremely powerful, but one that was also user-friendly. With this philosophy of high performance without offering up ease of use as a sacrificial lamb, Intel set about choosing the key components that make the 910 unique. The first and most obvious design choice is the half height, x8 PCI-E 2.0 form factor.

By opting for a half height form factor – but with full height PCI slot covers also included - the Intel 910 can fit into nearly any server, workstation or even consumer chassis on the market today. In order to fit all the NAND ICs, controllers and other components into such a small footprint, Intel had to opt for a three board layer cake style design instead of the typical dual board layout.

Naturally, the layout has increased cooling requirements over similar drives due to the lack of room between the hot running parts which causes less passive cooling potential via convection. As such a huge amount of airflow is needed to keep the Intel 910 within its operational limits. According to the documents we received, in standard configuration it needs an airflow rate 200 litres per minute and when in “high performance” mode requires 300LPM. As a side note, the high performance mode requires 38 watts peak and 28 watts average power be provided via the PCI-E slot as the 910 doesn’t use an external power connector.

Looking at the increase cooling and power specifications one would think that this ‘high performance’ mode would be similar to factory overclocking of a video card. This is actually not the case and a better analogy would be to refer to the Standard Mode is a factory under-clocked setting. By having this lower power, lower cooling mode as the default setting, the 910 is able to be used in a wide array of scenarios.

Cooling is a major consideration for enterprise administrators and the reduced heat – which does not sacrifice any of the specified performance – will be a welcome addition. By reducing the wattage requirements Intel was also able to ensure that the out of box configuration would fit within any PCI-E’s board specifications. However, as long as your system can handle the added power and cooling requirements, using the max performance mode does not void your warranty or in any way reduce the 910’s operational lifespan.

In practical terms, the high performance mode does require significantly more cooling and uses slightly more power, but neither should be of major concern outside of a server rack. The only potential stumbling point is making sure your case has optimal internal air flow as the 910 does run hot enough to cook itself without proper ventilation.

In order to keep from being able to fry eggs on the 910’s PCBs we simply orientated a pair of Noctua NF-P12-1300, 120mm fans to push air over and between the layers. This was more than enough to keep things cool, even in high performance mode and in all likelihood one fan would have been more than adequate.

It is also worth noting that you do not need to reboot to change between the two modes, a simple command via the command line is all that is needed. So if you do find the device overheating it is easy to rectify without any downtime, something which is just as important to Enterprise administrators as cooling and power consumption.

The actual hardware components Intel has selected also highlight the low hassle design philosophy of this model. While the Intel 910 is not bootable, it is virtually a “plug and play” device that will be recognized upon first bootup and simply listed as four separate “drives” – or Logical Unit Numbers (“LUN”) - in Windows’ Device Manager due to the RAID controller unit.

Unlike some other manufacturers, Intel opted for a single a LSI SAS2008 “Falcon” controller to facilitate communications between the 910’s connected albeit separate sections. The Falcon may not be precisely cutting edge, but it is a well respected controller known for its low latency, low CPU overhead and relatively high performance. More importantly, since it has been available –and widely used – for quite some time, most operating systems come with stock drivers already included. However Intel does recommend upgrading the drivers to their own in order to avoid issues.

Equally impressive is the fact that you the end user has complete control over the four LUNs and can configure them any way they please. If you require one, two, three or four “drives” the 910 is more than accommodating.

​

Further reinforcing this admirable ease of use philosophy, Intel has once again avoided the popular yet not very adaptable LSI/SandForce controllers for the four “drives”. Instead, they’re using the jointly developed EW29AA31AA1 Intel/Hitachi SSD chips.

These controllers are actually being used by Hitachi is using in their UltraStor solid state drives and unlike SandForce products they are highly capable of handling non-trim environments, which is what RAID and PCI-E based SSDs utilize by default. This choice of controller helps make the 910 a “hands off” drive which will not require constant monitoring like a SandForce based unit would. More to the point, the four EW29 controllers handle all of the “idle time garbage collection” at their level and require no additional CPU cycles to keep the 910’s NAND in tip-top shape.

Let’s begin by stating the obvious first: there are a dozen NAND slots that have gone unpopulated and they will be utilized for higher capacity models in the future.

Speaking of the NAND, unlike most enterprise devices this unit eschews SLC and standard MLC ICs for some simple reasons. Standard 25nm MLC NAND has an erase cycle life of between 3,000 and 5,000 which is very low for use in constant high demand environments like the typical server. By that same token the traditional choice of SLC - with its much longer life – would have significantly increased the upfront cost of an already expensive storage device.

Rather than opting for one of these two extremes, Intel has side stepped the issue by using what they call HET MLC NAND, or what the rest of the industry simply calls e-MLC or “enterprise MLC” NAND. High Endurance Technology MLC NAND is approximately thirty times more durable than standard MLC NAND but not nearly as costly to manufacture as SLC NAND. This in turn allows Intel to offer such massive drives for an actually reasonable price, while still being able to guarantee it for 5 years at 10 full drives writes everyday for those five years. To put that in more practical terms, Intel guarantees the drive’s NAND for over 14 petabytes of writes for the 800GB model and over 7 petabytes for the 400GB model.

The only negative to using such durable MLC NAND is that it does sacrifice some small file performance to gain such longevity. However, with four controllers working in tandem and such a dizzying number of NAND ICs, this ”issue” can easily be overlooked, especially considering the rated specifications of 2GB/s read, 1GB/s write sequential write performance and 180,000IOPS / 75,000IOPS small file read / write performance.

Also noteworthy are the four large capacitors this device comes equipped with. These are meant to provide more than enough reserve power to allow the 910 to flush its buffers and complete any outstanding writes in the event of unexpected power loss.

 

[SKYMTL]

ntel 910 and CPU Utilization

Intel 910 and CPU Utilization

Since the Intel 910 uses what is effectively a software RAID solution, higher CPU utilization is to be expected and we’ve seen plenty of similar solutions fall into this hole. While OCZ’s mass-market focused RevoDrives can get away with somewhat high CPU needs, any enterprise-oriented solution needs to keep overhead to a minimum in order to be remotely feasible for datacenter integration.

To obtain an accurate picture of exactly how much CPU horsepower this device truly requires we have configured the Intel 910 800GB device as a secondary “D” data drive and let the system idle for a few minutes. Once we aare satisfied an idle state has truly been reached, we used Windows’ built in Performance Monitor to see exactly how much processing power is being dedicated towards the storage solution. For comparison’s sake we have also included the results from the original RevoDrive 120GB, the RevoDrive3 x2 and finally with the system running a standard SATA 6GB/s Solid State drive as the “D” drive. Lastly we run Crystal DiskMark while monitoring CPU utilization.

Numbers this low are a testament to precisely how powerful the Intel 910 truly is. While RAID 0 has lower CPU overhead than most other RAID levels, this is the most typical configuration for such a device.

When compared against similar solutions, Intel’s device needs substantially less CPU overhead and you certainly don’t have to be worried about your server outputting excess heat and consuming more power just to provide processing for the 910. The only additional power and heat from using this storage solution will be from the Intel 910 itself which is something very few other devices of this caliber can offer.

 

[SKYMTL]

Testing Methodology

Testing Methodology

Testing a drive is not as simple as putting together a bunch of files, dragging them onto folder on the drive in Windows and using a stopwatch to time how long the transfer takes. Rather, there are factors such as read / write speed and data burst speed to take into account. There is also the SATA controller on your motherboard and how well it works with SSDs & HDDs to think about as well. For best results you really need a dedicated hardware RAID controller w/ dedicated RAM for drives to shine. Unfortunately, most people do not have the time, inclination or monetary funds to do this. For this reason our testbed will be a more standard motherboard with no mods or high end gear added to it. This is to help replicate what you the end user’s experience will be like.

Even when the hardware issues are taken care of the software itself will have a negative or positive impact on the results. As with the hardware end of things, to obtain the absolute best results you do need to tweak your OS setup; however, just like with the hardware solution most people are not going to do this. For this reason our standard OS setup is used. However, except for the Vista load test times we have done our best to eliminate this issue by having the drive tested as a secondary drive. With the main drive being a Phoneix Pro 120GB Solid State Drive.

For synthetic tests we used a combination of ATTO Disk Benchmark, HDTach, HD Tune, Crystal Disk Benchmark, IOMeter, AS-SSD and PCMark Vanatage.

For real world benchmarks we timed how long a single 10GB rar file took to copy to and then from the devices. We also used 10gb of small files (from 100kb to 200MB) with a total 12,000 files in 400 subfolders.


For all testing a Asus P8P67 Deluxe motherboard was used, running Windows 7 64bit Ultimate edition (or Vista for boot time test). All drives were tested using AHCI mode using Intel RST 10 drivers.

All tests were run 4 times and average results are represented.

In between each test suite runs (with the exception being IOMeter which was done after every run) the drives are cleaned with either HDDerase, SaniErase or OCZ SSDToolbox and then quick formatted to make sure that they were in optimum condition for the next test suite.


Steady-State Testing

While optimum condition performance is important, knowing exactly how a given device will perform after days, weeks and even months of usage is actually more important for most consumers. For home user and workstation consumers our Non-Trim performance test is more than good enough. Sadly it is not up to par for Enterprise Solid State Storage devices and these most demanding of consumers.

Enterprise administrators are more concerned with the realistic long term performance of any device rather than the brand new performance as down time for TCL is simply not an option. Even though an Enterprise device will have many techniques for obfuscating and alleviating a degraded state (eg Idle Time Garbage Collection, multiple controllers, etc) there does come a point where these techniques fail to counteract the negative results of long term usage in an obviously non-TRIM environment. The point at which the performance falls and then plateaus at a lower performance level is known as the “steady state” performance or as “degraded state” in the consumer arena.

To help all consumer gain a better understanding of how much performance degradation there is between “optimal” and “steady state” we have included not only optimal results but have rerun tests after first degrading a drive until it plateaus and reaches its steady state performance level. These tests are labelled as “Steady State” results and can be considered as such.

While the standard for steady state testing is actually 8 hours we feel this is not quiet pessimistic enough and have extended the pre-test run to a full ten hours before testing actually commences. The pre-test or “torture test” consists of our standard “NonTrim performance test” and as such to quickly induce a steady state we ran ten hours of IOMeter set to 100% random, 100% write, 4k size chunks of data at a 64 queue depth across the entire array’s capacity. At the end of this test, the IOMeter file is deleted and the device was then tested using a given test sections’ unique configuration.


Processor: Core i5 2500
Motherboard: Asus P8P67 Deluxe
Memory: 8GB Corsair Vengeance LP “blue”
Graphics card: Asus 5550 passive
Hard Drive: Intel 520 240GB
Power Supply: XFX 850

 

[SKYMTL]

ATTO Disk Benchmark

ATTO Disk Benchmark

The ATTO disk benchmark tests the drives read and write speeds using gradually larger size files. For these tests, the ATTO program was set to run from its smallest to largest value (.5KB to 8192KB) and the total length was set to 256MB. The test program then spits out an extrapolated performance figure in megabytes per second.

​

As expected the Intel 910 does scale very well with larger sized data, but on the low end the difference between it and the OCZ RevoDrive is rather minor. This is pretty much to be expected as “software” raid solutions are not tweaked or tuned for small file sizes and the additional overhead from the RAID aspect does negate some of the advantages from using four controllers. Of course this test suite does not exactly play to this device’s strong suit as it uses extremely low queue depth tests by default whereas the Intel 910 has been fine tuned for deep queue depths which are more typical of server environments.

 

[SKYMTL]

Crystal DiskMark

Crystal DiskMark

Crystal DiskMark is designed to quickly test the performance of your hard drives. Currently, the program allows to measure sequential and random read/write speeds; and allows you to set the number of tests iterations to run. We left the number of tests at 5 and size at 100MB.

The large and mid-sized results are simply jaw dropping when the Intel 910 is fully utilized in a four LUN / drive RAID array butthe performance does not precisely scale perfectly. This is once again due to the software RAID nature of this device.

AS-SSD

AS-SSD is designed to quickly test the performance of your drives. Currently, the program allows to measure sequential and small 4K read/write speeds as well as 4K file speed at a queue depth of 6. While its primary goal is to accurately test Solid State Drives, it does equally well on all storage mediums it just takes longer to run each test as each test reads or writes 1GB of data.

As with Crystal Diskmark, the large and mid size numbers the Intel 910 post are indeed good and the Maximum Mode's performance boost is downright impressive.

 

[SKYMTL]

IOMETER: Our Standard Test

IOMETER: Our Standard Test

The goal of our IOMeter File Server configuration is to help reproduce a typical multi-user storage server. As with most storage servers, the majority of the typical file server’s workload is highly random in nature with a majority of data requests being reads with a few writes of various sizes.

To test each drive in such a scenario we ran 6 test runs per device (1,4,16,64,128,256 queue depth) each test having 8 parts, each part lasting 10 min w/ an additional 20 second ramp up. The 6 subparts were set to run 100% random, 75% read 25% write; testing 512b, 4k,8k,16k,32k,64k size chunks of data. When each test is finished IOMeter spits out a report, in that reports each of the 6 subtests are given a score in I/Os per second. We then take these 8 numbers add them together and divide by 6. This gives us an average score for that particular queue depth that is heavily weighted for file server usage.

In the first chart we have used our standard 1 client with 1 worker. In our second chart we have used two clients, each with two workers or four times the concurrent operations at a given queue depth.

There is no denying the power this PCI-E based device has but it only pulls far ahead of a OCZ RevoDrive3 x2 480GB – a much less expensive drive– when the queue depths are deep. You see, Intel's 910 has been optimized for the Enterprise market where deep queue depth performance matters more than shallow depth performance.

On the positive side, the e-MLC NAND is not proving to be much of hindrance in workstation scenarios and the two client, with two workers results are simply astronomical. It is also interesting to see the boost which the "high performance” mode bestows upon an already extremely quick SSD.

 

[SKYMTL]

IOMETER: File Server Test

IOMETER: File Server Test

To test each drive we ran 6 test runs per device (1,4,16,64,128,256 queue depth) each test having 8 parts, each part lasting 10 min w/ an additional 20 second ramp up. The 6 subparts were set to run 100% random, 75% read 25% write; testing 512b, 4k,8k,16k,32k,64k size chunks of data. When each test is finished IOMeter spits out a report, in that reports each of the 6 subtests are given a score in I/Os per second. We then take these 8 numbers add them together and divide by 6. This gives us an average score for that particular queue depth that is heavily weighted for file server usage.

In the first chart we have used our standard 1 client with 1 worker. In our second chart we have used two clients, each with two workers or four times the concurrent operations at a given queue depth.

​

Since the File Server configuration is more analogous to our standard IOMeter configuration, the results are unsurprisingly very similar and do not tell us anything new. Namely, the Intel 910 is designed for situations and scenarios where the queue depths are so deep that a typical storage devices would be overwhelmed. We do have to wonder how much potential is being lost because it is relying upon a “software” RAID configuration, but the greatly reduced CPU overhead more than makes up for this potential negative.

 

[SKYMTL]

IOMETER: Web Server Test

IOMETER: Web Server Test

The goal of our IOMeter Web Server configuration is to help reproduce a typical heavily accessed web server. The majority of the typical web server’s workload consists of dealing with random small file size read requests.

To replicate such an environment we ran 6 test runs per device (1,4,16,64,128,256 queue depth) each test having 8 parts, each part lasting 10 min w/ an additional 20 second ramp up. The 8 subparts were set to run 100% random, 95% read 5% write; testing 512b, 1k, 2k,4k,8k,16k,32k,64k size chunks of data. When each test is finished IOMeter spits out a report, in that reports each of the 8 subtests are given a score in I/Os per second. We then take these 8 numbers add them together and divide by 8. This gives us an average score for that particular queue depth that is heavily weighted for web server environments.

In the first chart we have used our standard 1 client with 1 worker. In our second chart we have used two clients, each with two workers or four times the concurrent operations at a given queue depth.

The Intel 910 has phenomenally impressive read performance and the four RAID array-based results are simply stunning. Unlike the other IOMeter results, the HET NAND’s write performance handicap is a moot issue here. Unfortunately, the shallow queue performance is still a touch lower than we would like to see.

 

[SKYMTL]

IOMETER: Email Server Test

IOMETER: Email Server Test

The goal of our IOMeter Email Server configuration is to help reproduce a typical corporate email server. Unlike most servers, the typical email server’s workload is split evenly between random small file size read and write requests.

To replicate such an environment we ran 5 test runs per drive (1,4,16,64,128 queue depth) each test having 3 parts, each part lasting 10 min w/ an additional 20 second ramp up. The 3 subparts were set to run 100% random, 50% read 50% write; testing 2k,4k,8k, size chunks of data. When each test is finished IOMeter spits out a report, in that reports each of the subtests are given a score in I/Os per second. We then take these numbers add them together and divide by 3. This gives us an average score for that particular queue depth that is heavily weighted for email server environments.

In the first chart we have used our standard 1 client with 1 worker. In our second chart we have used two clients, each with two workers or four times the concurrent operations at a given queue depth.

Thanks to the unique nature of both the test and Intel 910’s architecture, the results posted are merely good at the lower queue depths. It is obvious that Intel has designed this device with deep queue depths in mind and while this will somewhat handicap its workstation appeal, the 910 does cater to the Enterprise niche where deep queue depths and durability are core requirements. So as the queue depths get deeper this handicap is quickly obfuscated.

It is also bears mentioning again that some of this handicap does come from the HET / e-MLC NAND. However, a small performance hit in non mission critical areas areas in return for greater NAND durability is something that will greatly appeal to Enterprise IT administrators.