Search Results Category:

USB 3.0 tests – LaCie and Asus cards comparison

by eduardo8088
Published on: April 11, 2011
Categories: Disk Drive tests, USB 3.0
Tags:No Tags
Comments: No Comments


We conducted some test on USB 3.0 and as always, we test in order to have the real numbers -we don’t trust manufacturer specifications-.

We are going to compare two different models of USB 3.0 cards using a LaCie RAID external drive so we can have a top speeder disk that is no a limiting factor in tests.

The testing program is Crystal Disk Mark 3.0.0 x64. I did 29 replicas of a 2 GB test for each card and compared them using a non parametric statistical test and plots of data points.

Our test setup is:
Processor: Dual Xeon E5420, 2,5 Ghz, 12 MB cache, 1333 FSB
System drive: Seagate ES.2 ST3500320NS, 7200 rpm, 500 GB
Video card:    NVidia GeForce GT430, 1GB
Board:    Server board, dual socket, Intel 5100 chipset
PSU:    Real 500W high efficiency HEC
Operating System:    Windows 7 Ultimate 64bit, English
Testing Drive: LaCie 2big USB 3.0 Drive, RAID0 2 hard disks, 2 TB total external drive
Testing card1: LaCie PCIe x4, USB 3.0. 2 port adapter, NEC µPD720200 chip
Testing card2: Asus U3S6 MG990T-C PCI-E x4, USB 3.0 2 ports + SATA 6Gb/s 2 ports

Read comparison USB 3 plot














Figure 1. Reading transfer speed of two USB 3.0 cards in 30 repicas of a 2000 MB test. Error bars are Standard Mean Error

Figure 2. Writing transfer speed of two USB 3.0 cards in 29 replicas of a 2000 MB test. Error bars are Standard Mean Error


To test if there are significant differences between the two cards I choose the Mann-Whitney non-parametric test (α 0.05000, N=29).

For reading speed , there are significant differences (p-level=0.00001, U=137, reject that both samples come from the same distribution). Knowing that both card are different, a simple inspection of the figure 1 reveals that LaCie card have more performance:  The area bellow the LaCie card line (blue line) is greater than the Asus one (red line).

For writing, a simple inspection of figure 2 makes the statistical test almost unnecessary and is evident which card is better, accordingly, I found significant differences (p-level=0.00000, U=0) between two cards and again, LaCie card performs better.



LaCie card performs better than Asus card, i suspect that partly is due the Asus card also have a SATA 6 Gb/s controller, adding a internal steps managing bandwidth.

By the way, the mean read transfer rate is  226.004 MB/s and the mean write transfer rate is 154.186 MB/s in the LaCie. Despite that number is valid only at the beginning of the drive, is pretty much similar to a eSATA drive transfer rate, making USB 3.0 a very good substitute to eSATA with the advantage that USB 3.0 drives are cheaper that eSATA drives and you can plug them in a USB 2.0 port if needed. We should conduct a test to probe this appreciation.






CUDA performance SETI@home

by eduardo8088
Published on: April 8, 2011
Categories: CUDA tests
Tags:No Tags
Comments: No Comments

This is just a simple comparison of performance on SETI@Home between two of our computers: The 4 Nvidia Tesla powered supercomputer ( and a small 2 Quad Xeon 5420 +Nvidia 8500GT.

SETI@Home is a distributed computing project looking for extra terrestrial intelligence analyzing radio signals recorded at Arecibo radio telescope, distributing that analysis in home computers over Internet. More info at:

Graph SETI@Home performance 8 cores 1 GPU







Figure 1. Credits obtained during 15 days on a dual Quad Xeon 5420 + 1 NVidia 8500GT

Graph SETI@Home performance 16 threads 4 GPU

Figure 2. Credits obtained during 15 days on a dual Xeon 5620 + 4 NVidia Tesla C1060

SETI@Home gives you credits for each processing task completed and validated. The graphs above show how much credits where assigned to each machine during a two week time frame. Between 20th Aug. and 6th Sep. SETI were having  problems with their servers and task downloading was slow, despite that, Tesla machine accumulated more than double the credits than Quad Xeon machine. After 6th Sep., both machines downloaded task normally and the difference is huge, the Tesla machine processed a lot more than the dual Quad Xeon. Unfortunately, the test where  stopped after 15 days because electricity bill due the 850W drawn by Tesla computer at almost 100%.



IOps comparison: Solid State Drive (SSD) vs. 15000 vs. 7200 vs. 5900 vs. 5600 rpm

by eduardo8088
Published on: January 24, 2011
Tags:No Tags
Comments: No Comments

In this test we are going to compare the input/output operations per second (IOps) between several hard drives we have by hand: Two enterprise 15000 rpm SAS (Serial Attached SCSI) hard drives, three high performance desktop 7200 rpm SATA drives, two low end desktop SATA drives and an enterprise SATA solid state drive.

To go directly to the comparison chart, click here.

Comparing hard drives is not only important in how many MB/s they are capable to transfer, it is important too to know how many IOps they are able to handle, that is, how many times per second they are able to read or write data in a given time. The IOps is a very important parameter in machines that handle lots of requests, e. g. computers running a database, which typically read and write a huge amount of small packets of data in a short amount of time.

With an every day example would be possible to easily understand the importance of  IOps. Think in that you are going to copy a 2 GB AVI movie and then you are going to copy a 2 GB folder of pictures. The movie will copy at maximum speed, for say 100 MB/s and will take 20.5 seconds. The pictures will copy at 12.5 MB/s and will take almost 3 minutes. You can try this at home with similar files and you will see the huge difference  among copying one thing and the other.

¿What is the difference in the example above if both are the same size? In simple (and coarse) terms, to copy the movie the hard disk head only needs to read the file position and then copy its content, only two operation are needed. To copy the pictures, the hard drive head needs to read the position of every single picture and copy its content, one by one; if each picture size is about 1.5 MB, there will be by 1500 pictures, that means by 3000 operations to copy all the pictures. In each operation the head takes a certain amount of time to finish the task and that causes a drop in the  read or write speed of the disk.

With the developing of faster,bigger and reliable non-volatile memory chips (that doesn’t loose its content after cutting the power), new storages devices appeared in the market using the interface of the standard hard drives but instead of platters they used memory chips. Called Solid Stated Drives, a couple of years ago they had sky level prices (several thousands of dollars) but speeds that surpassed by far their mechanical counterparts produced  increased sales and ultimately prices decreased with time.

The first models had problems with lifetime of memory chip sectors, that had a limited amount of read/write cycles, but nowdays internal SSD  controllers have write leveling algorithms that keeps sector usage levered across the chips, equaling or surpasing MTBFs of disk drives. Additionally prices are getting lower and speeds keep increasing in each new model.

SSDs surpasses their disk counterparts in transfer rate (MB/s) twofold or more, however,  the IOPS results are even more astonishing. In order to have concrete numbers of how many IOps a SSD is capable compared to a disk drive, we conducted a test using IOmeter usig a configuration profile posted in the Communiti forums of VMWare (, so results could be comparable with others.


The test simulatd a real life access to a database, with a transfer request size of 8 kB, 35/65% of random/sequential distribution; 65%/35% of read/write distribution and 64 outstanding IOps. The test ran for 5 minutes and was repeated 6 times per disk. The chart reports the mean and the standard error of the mean for each disk.

Additionally, the test were conducted over the last 10% of each drive because it is the slowest area in the disks, that means that it is a limiting factor of performance hence a representative (an more real) parameter of  the behavior of the disks in a critical condition, that is, with the drives almost full.

Drives to test:

Seagate Cheetah 15k.5 ST373455SS, SAS 3Gb/s, 15000 rpm, 16 MB cache, 74 GB.
2. Fujitsu MAX3147RC, SAS 3Gb/s, 15000 rpm, 16 MB cache, 146 GB.
3. Seagate Cheetah 10T ST3300555SS, SAS 3Gb/s, 15000 rpm, 16 MB cache, 300 GB.
4. Seagate Barracuda ES.2 ST31000640SS, SATA 3 Gb/s, 7200 rpm, 16 MB cache, 1TB.
5. Seagate Barracuda ES.2 ST3500320NS, SATA 3Gb/s, 7200 rpm, 32MB cache, 500GB.
6. Seagate Barracuda 7200.12 ST31000528AS, SATA 3Gb/s, 7200 rpm, 32MB cache, 1TB.
7. Seagate Barracuda LP ST32000542AS, SATA 3Gb/s, 5900 rpm, 32MB cache, 2TB.
8. Hitachi  Deskstar 5K1000 HDS5C1010CLA382, SATA 3 Gb/s,  5600 rpm, 8 MB cache, 1 TB.
9. SuperTalent Technology TeraDrive CT  FTM12CT25H , SATA 3Gb/s, SSD, 120G.

The three first are enterprise SAS disk drides, fourth and fifth are SATA enterprise disk drives, sixth es a desktop performance SATA disk drive, seventh and eighth are home desktop disk drives and the last is an entrerprise Solid State Drive.

System Setup:

-2 x Processor Intel® Xeon E5620 ; 2,40 GHz, 12 MB cache, 1333FSB,  Intel® HyperThreading technology, Intel® Turbo Boost technology.
-24 GB RAM DDR3 1333MHz ECC, (4GB x 6), triple channel config.
-SSD  (Solid State Drive), 80 GB Intel X25M SSDSA2M080G2GC for the Operating System.
-SAS/SATA 6Gb/s controller, PCI-E 2.0 x8, HighPoint RocketRaid 2720, 8 ports.
-Video card NVidia GeForce 7300 GT.
-Mother board Supermicro with Chipset dual  Intel® 5520, 36 líneas PCI-E 2.0 cada uno (72 total), 6.40 GT/s cada uno.
-2 x Gigabit Intel® 82574L LAN.
-Chasis 4U tower/rack muontable, high efficiency 1400W PSU.
-Windows 7 Ultimate de 64bit.

Well, lets the data chart talks by itself:

Mean IOps Comparison Chart
Mean IOps Comparison Chart

The difference between disk drives and the SSD was so big that I put labels in th chart to improve readiness.

Despite this is not a rigorous statistical test,  it is enough to illustrate that differences are abysmal.   The SSD is 18.5 to 21.6 time faster than a 15000 rpm SAS disk drive, but both have similar prices. Just imagine a RAID with this SSDs!

The big difference in IOps between drives is because the SSD controller needs very few time to locate a memory address in the chip, so while the hard drive head executes an operation the SSD executes 20 operations, that is, the disk head needs to fly over the disk until right position but the SSD simply locates electronically a row and a column to address the memory chip, a much faster operation.

This technology is in active development so we can expect a increase of perfromance of SSDs in the years coming, particularly because the interface bandwidth is increasing too. We already have SAS/SATA of 6 Gb/s and the 9 Gb/s is coming; plenty  of transfer rate room.

We also need to wait what is the answer of hard drive manufacturers. The started with hybrid hard drives, but such a big differences and prices of SSDs falling is difficult to think that disk drives are going to keep the lead in the market for long time.

As a tip, all of our new workstations for renting  are being produced using the SSD of this test as an operating system drive, so Windows and applications response times improved notably. We also have a customer whit a productions database implemented over 3 RAID1 of SSDs, running successfully and beating their previous SAS 15k system.

Stat tuned for more tests!




CUDA multi GPU TESLA super server specifications

by eduardo8088
Published on: October 25, 2010
Categories: CUDA tests
Tags:, ,
Comments: No Comments

I wish to show you all a picture and the specs of our testing CUDA multi GPU TESLA super computer (click to enlarge): TESLA GPU Supercomputer pic

The four big prominent cards are four NVidia TESLA C1060 GPU Computing Processors, that could be considered blind video cards with lots of RAM that are used for processing data, not for video games; also you can see two heatsinks under each one are a Intel® Xeon E5620 processor (2,40 GHz, 12 MB cache, 1333FSB), 4 cores, 8 threads each one, 16 threads total; finally there are four big hot-swappable system fans; everything in a Tower chassis 4U rackmountable.

This machine was the one  used in our previous post: CUDA Tests – GPU render – Bunkspeed Shot

The (almost) full specs are:

4 x NVidia Tesla C1060 Computing Processor, 4 GDDR3 RAM, 240 Streaming Processor Cores each one (960 cores total).

2 x Intel® Xeon E5620 Processors; 2,40 GHz, 12 MB cache, 1333FSB, simultaneous multithreading (Intel® HyperThreading Technology), Intel® Turbo Boost Technology.

24 GB RAM DDR3 1333MHz ECC, (4GB x 6), tripple channel configuration, up to 192 GB RAM.

100 GB SSD (Solid State Drive) OCZ Vertex Limited Edition, up to 270 MB/s read, up to 250 MB/s write (235 MB/s sustained write).

NVidia Quadro FX 580 video card, 512 MB GDDR3 RAM.

Motherboard with dual Chipset Intel® 5520, 36 PCI-E 2.0 lanes each one (72 total), 6.40 GT/s each one.

2 x Intel® 82574L Gigabit Ethernet ports.

Chassis 4U rackmountable/ tower, 8 x 3,5″ remobable HDD bays, 3 x 5.25″ bays, 1 x 3.5″ bay, 4 x hot-swap internal fans, 2 x hot-swap rear fans, redundant 1400W high-efficiency power supply.



CUDA Tests – GPU TESLA render – Bunkspeed Shot

by eduardo8088
Published on: October 22, 2010
Categories: CUDA tests
Tags:, ,
Comments: No Comments


In the last two months we started to test how a technology based in GPU porcessing works. It’s named CUDA and is developed by NVidia.

NVidia have cards dedicated only to GPU processing, TESLA cards, so we decided to buy a computer with four of them to see how it works.  We named the computer “Tesla”  (yes, not original but very handy for everyday use) and this is one of the first test that we made with the machine.

The configuration is as follows: 4 NVidia Tesla C1060 4GB GPU cards, 2 Xeon E5620 processors (16 threads total), 24 GB DDR3 ECC RAM Tripple channel, 100 GB OCZ Vertex SSD, 1400W Redundant Power Supply, Tower chassis, Win 7 Ultimate 64 bit.

Since we are mainly dedicated to the video 3D and post-production industry, we choose to do several tests using the GPU rendering applications in the market. We started with Bunkspeed Shot, an standalone application based in iray, the GPU render engine of Mental Images, the creators of the well known render engine Mentalray.

In the test, we just played around with a sample model of the program, manipulating it in the viewport and then changing the materials to see how it behaves. Finally, we changed the viewport size to fit the screen (1280×720) trying to get more of the Tesla cards. The performance looks very good with the 4 Tesla, however, the software didn’t use all the capacity available of both, the CPUs and the GPUs. We are trying to find how to bring the GPUs and CPUs near to the 100%, so hang with us to see what we find!

Just watch the video bellow to judge by yourself how the test went (don’t forget to switch to HD quality to see every detail of the test!).




There is a new post with a picture and the detailed specs of the machine, take a look:

Welcome to the blog

by admin
Published on: May 30, 2010
Categories: General
Tags:No Tags
Comments: No Comments


In this blog we are going to post all our benchmarks and tests of our research in the latests hardware and software we use in our company.
The main idea is to show how our machines behave in the real word, the real numbers behind the specifications.
Feel free to ask all general questions on this post.

Stay tuned!

page 1 of 1
Welcome , Today is Sunday September 20, 2020
mulberry outlet coach outlet burberry outlet coach factory outlet mulberry outlet coach outlet UGG Pas Cher cheap oakley sunglasses cheap nfl jerseys wholesale nfl jerseys coach outlet canada black friday coach ugg boots on sale cheap uggs gucci outlet oakley outlet coach outlet