SGI CEO Blog

Update on Japan

To our employees, customers and partners: 

Following our recent acquisition of SGI Japan, Ltd. on March 9, we have received many questions and concerns regarding the effects of the earthquake and tsunami that affe cted the area.

I was in Japan meeting with our new employees when the earthquake and the resulting tsunami struck last week. The earthquake was severe and the resulting devastation, especially in the northeast part of the country, are heartbreaking. Japan is the most prepared nation in the world for these types of disasters, and even with the best engineering, planning and education, these events show us all humility. 

As we are all witnessing, the situation in Japan is evolving and has not yet stabilized.

Our deepest condolences go out to the people of Japan and to those affected by this tragedy. 

I have been touched by the numerous inquiries from employees and customers offering to help those affected by the disaster. SGI has set up a site where the SGI community can contribute to help bring aid to the people of Japan. Every dollar helps.

Our immediate priority is for the personal safety and well being of our employees as well as the needs of our customers. We are encouraged with the initial reports that all of our employees are safe and accounted for and we have emphasized that each employee should do what is best for their family based on their individual circumstance – and that safety and well-being takes priority.

The resolve, calm and spirit of the team in Japan are beyond what words can describe. Despite extreme hardship and the many challenges directly ahead, the SGI Japan team is also focused on helping our customers. The team is standing firm and ready to do what it takes to help.

Last week’s earthquake wasn’t the only major quake I’ve experienced firsthand. I was ten miles from the epicenter of the Loma Prieta quake, which struck the Bay Area in 1989. While the devastation was not on the same scale as the earthquake and tsunami that hit Japan, the damage was also widespread, infrastructure was destroyed and many towns were cut off.  It had an adverse impact on many people. I will never forget those ensuing days of aftershocks, grief, isolation, uncertainty, and waiting in long lines just to receive basic services. I equally remember the simple joy of being able to reconnect with family, friends and colleagues in the following days.
 
It is not possible to fully comprehend the unfolding events in Japan and life there will never be the same.  But people come together, adjust, rebuild, and they emerge stronger with a more collective sense of purpose.

 
 

The Prism and the Pendulum

For the first time in the history of computing, an accelerator-based architecture is powering the world’s fastest machine. That machine is the Tianhe-1A, a bespoke machine (custom one off) that is not commercially available (Tianhe translates literally into “sky river” or the Milky Way.)

SGI today introduced SGI Prism™ XL, a purpose-built insanely-scalable accelerator platform, based on standard components and open software, leveraging a new architecture named SGI^® STIX™. 

The difference between Prism XL and Tianhe-1 is that Prism XL can scale to 20 times that of Tianhe-1 and is a commercially available product, built on standard technologies, using open software.

Up to 20X the scale. Open. Standards Based. Commercially Available. 

It is simple arithmetic. A 125 cabinet deployment of Prism XL could potentially replace the entire TOP 100 fastest computers in the world.

The Double Precision scale for Prism XL looks like this:

1 Cabinet	.2 PFLOPs
5 Cabinets	1 PFLOPs
50 Cabinets	10 PFLOPs
100 Cabinets	20 PFLOPs
200 Cabinets	40 PFLOPs

Today’s announcement also completes the SGI architectures:

1. Scale-out: with Rackable and Altix ICE
2. Scale-up: with Altix UV
3. Accelerators: with SGI Prism XL

This is a significant step towards ExaScale and moves the pendulum for technical computing.

SGI Prism XL is an open, standards-based accelerator platform for large-scale computation, supporting single and double-wide GPU cards from nVidia and AMD/ATI, as well as Tilera Tile-based processing cards – both leveraging the full bandwidth of PCIe interconnect and on-board dual QDR Infiniband or GigE (10Ge or 40Ge) for cluster integration. Each STIX can also have local direct attached storage (SATA and SSD).

Here is a picture of a STIX:

SGI Prism XL is a long-life platform and will easily adapt as technology advances. As more PCIe accelerator-based technologies come to market, plug them in. As PCIe Gen 3 comes to market, PRISM XL will support it. As GPGPU’s gain smarter interconnects, we will support that. As probability processors come to market, plug them into Prism XL.

Another great example is that Prism XL will support the Knights Ferry software development platform for Intel® Many Integrated Core Architecture.

Furthermore, given our cluster interconnect of GigE and Infiniband, many topologies will be possible: from fat-trees to tori to hypercubes.

Also, our new architecture and packaging were designed with extreme density and scale in mind.  To achieve density we needed to go “beyond the blade” – the basic building block is not a chassis, not a blade, not a tray – but a “stick” – or SGI STIX whose dimensions are 5 x 3 x 37 inches. For example, an x86 PFLOP takes roughly 100 cabinets today, but with our new architecture, it takes only 5 cabinets and reduces the facilities footprint from a city block down to a small car. 

As you need to scale, simply add a STIX at a time.

We see strong advantages for GPU accelerators in oil & gas for seismic analysis, image and signal processing within defense & intelligence, rendering and image processing within media & entertainment, as well as bio-informatics and physics.

I would like to highlight that accelerators are not just GPUs. For example, Tilera is NOT a GPU; it is a tile-based processor (think fast integer arithmetic). Tiles are low power and ideal for encrypt/decrypt, image and signal processing, deep packet inspection, format conversion and certain cloud-based content delivery.  SGI Prism XL supports Tilera. We also support Violin and Fusion IO. Prism XL also supports high-speed I/O cards for high-speed data ingest and data acquisition. Below is a picture of what a fully populated cabinet would look like:

Equally important to the hardware is the software environment.Prism XL supports the SGI run-time and programming tools for massively parallel-environments that we support on our scale-out and scale-up lines. This includes standard LINUX, CUDA and OpenCL compiler tools, SGI Management Suite for provisioning to production, as well as job scheduling packages.

SGI Prism XL general availability is scheduled for December 2010.

Our customers now have 3 choices of technical computing architectures:

Scale-out:	Rackable and Altix ICE
Scale-up:	Altix UV
Accelerators:	Prism XL

This is the Prism and the Pendulum.

Visual Computing

SGI is focused on 5 technology platforms: scale up, scale out, storage, software, and data center technologies. Our services organization deploys these platforms, partners with relevant third parties, and fulfills customer requirements with their expertise. Our technology platforms and services together ensure we are both relevant and complete.

SGI's mission focuses our engineering at the extremes of scale, speed, power, density, and latency. SGI customers use our solutions for many diverse applications: strategic science, national defense, modeling climate change, designing safer, more energy efficient cars, powering the Internet, to visual computing.

Wind the clock back 20 years. Remember when the floating-point processor was not integrated with the integer processor? Motherboards had two sockets: one for the 386 or 486, and you purchased and plugged in (ever so gently) a math co-processor. Remember the Weitek 4167? The 4167 offered better performance than an integrated FPU in the 80487. What enabled the 4167 to outperform others was its chip technology and specialized compilers that generated specific and optimized instruction sets for the dedicated FPU. This source of innovation was driven from the scientific community, based on the need for speed!

This architectural approach of separating integer and floating point arithmetic has recently reemerged with the GPU. The GPU is not just for gamers anymore. SGI believes there is an enormous customer benefit to incorporating the GPU into its core platforms for both computational acceleration and for visualization. This is why we are now focusing more R&D investment into advancing GPU products by means of its integration into our core platforms and applying these products to real-world problems.

Graphics processing units are proven technology in the world of gaming. These processor crunch models, numbers and display results faster than a traditional CPU. This technology has emerged in the world of high performance computing. GPUs have become important - and as software tools and the ecosystem mature (languages, compilers, libraries, debuggers, profilers, schedulers) - could even emerge as general purpose (the GP-GPU). Several applications for molecular dynamics, material science, CFD, seismic, astrophysics climate, medical, etc. have been ported or are being ported to GPU. I expect that in the next couple years we'll see applications running on clusters with GPUs in each node.

SGI has a long and successful history of helping our customers advance at the key junctures of "technology shift." We have helped customers migrate from custom MIPS-based CPUs to Itanium and then to 64-bit x86 processing, from UNIX/IRIX to standard Linux, and now, to GPU technology.

Visual computing can be thought of as (1) receiving lots of data from many sources, (2) storing that data or processing it in real-time, (3) processing the data against models, (4) then displaying the result. The last few years have seen model sizes explode in size, growing at a rate faster than Moore's law. Data feeds are growing and input pipes are getting wider and more diverse. Our HPC compute and storage solutions enabled the storing of data and its models as wells as the processing and displaying of results.

Customers frequently have many varied and complex display requirements that require deep graphics competency to implement. Our services organization draws on many years of experience and highly competent staff in putting together solutions of our customers' needs. Using a combination of our own technology coupled with best of breed third party packages, we put together and guarantee integrated solutions and results for our customers. Our customers have depended, and can continue to depend on SGI to provide leading edge systems aimed at collaboration, remote visualization, and decision support.

Our vision is to provide an end-to-end visual computing solution through our core platform technology and expert services. The building block for our scalable platforms contains GPU as a core component. We recently expanded our support with the introduction of the CloudRack X2 Workgroup Cluster incorporating a wider set of options from nVidia and ATI/AMD. It is an ideal visualization and strategic science  platform. As customers scale up and scale out for large workload needs, we offer integrated Nehalem and GPU technology. Further, our next generation platform, Ultraviolet, will also support GPUs.

The future is exciting.

Let's bring this to life with a couple of real world examples.

Reverse Time Migration (RTM) is a technique for generating images of subsurface geological structures. RTM is core to seismic imaging, a kind of earth echography. Floating point speed is more important than getting every pixel just right. The GPU can offer RTM techniques close to 200 GFlops.

Another example is heart simulation to discover electrical triggers for early disease diagnosis and treatment. The rendering of the pixels is not the bottleneck, but rather, taking terascales of data and filtering and mapping to relevant information sets prior to visualization. In this example, big data visual computing is a data reduction exercise. In 1980, there were 100K of triangles from a server to 1M triangles/pixels to a user. Over the last 30 years, the 100K has grown to 1TB of data, though the triangles/pixels to a user has remained constant.

SGI HPC compute and storage technology, integrated with GPU technology, is being used for the exploration of natural resources and to save lives through medical research. I need no further examples than these to understand where our R&D resources are best placed.

This is why we are focused on working closely with nVidia and ATI/AMD and Intel many-core Larrabee, integrating their advanced graphics technology into our core platforms versus writing software to replace the GPU. When speed truly matters, put it in hardware.

Shift Happens. And the shift is the GPU.

The Workgroup Cluster

London is one of my favorite cities in the world, for it is truly cosmopolitan. Where else in the same day can you both rub the Rosetta Stone and have a cappuccino in a crypt? Now, I am not a "subway" kind of guy, but I enjoy traveling the London "Tube." The London Underground Authority has a real flair for style, from their tube layout diagrams, chic logos, and memorable recorded expressions. One such expression that comes to mind is "Mind the Gap."

CloudRack^TM X2 minds the gap. That is, the gap between the desk-side workstation (8 cores) and the fully populated rack-level cluster (800 cores). The CloudRack X2 fills this gap and we are pleased to introduce a true workgroup cluster.

The CloudRack X2 scales up to 216 cores and can be used free standing or in a 19 inch data center rack. It is a tray-based design offering configurations for Internet and Cloud Computing, High Performance Computing, and Graphics Clusters - all within the same workgroup configuration - with high speed interconnect, processor and graphics/GP-GPUs.

We also leveraged our Power XE^TM technology from CloudRack C2 for 12VDC power distribution allowing for an overall efficiency of up to 93 percent. Further, our autonomic fan bank arrays allow customers to run at higher ambient temperatures, and in this case, up to 40 degrees Celsius (104 degrees Fahrenheit). See my prior post, "Some Like it Hot." Power XE also allows for N+1 redundancy. Lastly, the AC total harmonic distortion is below 5 percent with no single harmonic above 3 percent. Simply said, run more servers on the same circuit.

On casters, its small form-factor makes it ideal for engineering and laboratory environments. Alternatively, up to three 14U CloudRack X2 enclosures can be easily rack-mounted in any industry-standard 19 inch rack, scaling to the largest of data center deployments.

The CloudRack X2 is our first product launch as SGI. To that end, we have launched with a wide range of socket options: Intel 5500 series processors (Nehalem), AMD 2300 or 2400 processors (Shanghai or Istanbul), Intel Atom, AMD Phenom X4 and Athlon X2 and NVIDIA and onboard InfiniBand support for some configurations.

CloudRack X2 supports Microsoft Windows Server 2008 and Microsoft Windows HPC Server 2008, Linux, and it is Intel Cluster Ready.

Mind the Gap!  

Read My FLOPS

There have been a couple of articles on the SGI view of Itanium and Shared Memory over the last few days: one from HPC Wire; and the other from eWeek. The articles are well written and worth reading. There is, however, a basic point that needs a bit more emphasis.

SGI is 100% committed to Itanium.

SGI Altix Itanium systems power some of the most important systems in the world - from conducting global warming research to fraud detection to homeland security. Itanium offers unique value in RAS(reliability, availability, serviceability) and memory addressability unmatched by any other non-strictly-proprietary microprocessor. We expect to continue to sell SGI Altix Itanium for many years to come.

Another example of our commitment is Altix Itanium at the Geophysical Fluid Dynamics Laboratory at NOAA. The GFDL mission is "to development computer simulations to improve the understanding and prediction of the behaviors of the atmosphere, the oceans, and the climate." SGI is honored to be part of this mission. GFDL has a wonderful presentationhighlighting SGI Altix Itanium at work and the complex problems the architecture solved. It is highly relevant work to today's sensitivities.

Altix is a large shared-memory, single system image (SSI) environment. That is to say, we run standard Linux, and applications running in the OS can see ALL memory across ALL the compute nodes. SPEC CPU2006 rate benchmarks are the standard test. The SPECfp_rate_base2006 world record is on an Altix with 1024 Itanium2 cores, which clocked x5.3 faster than the next pure shared-memory competitor. Fast and relevant is a good place to be.

But SGI will not rest on its success. Currently, we have an internal project, code-named "Ultraviolet." It represents the next generation of Altix, designed to utilize both Itanium and Xeon processors. Our first priority is to develop a Xeon version of Ultraviolet, based on the strong feedback we have received from many customers.

We continue to develop and expand our Altix Itanium line of products, and Ultraviolet V1 is focused on Xeon.

So, as we progress down our shared memory roadmap, we will continue to evaluate our processor choices with our customers, keeping our eyes open in terms of the many standard processor options available in the market. As Intel continues to deliver its QuickPath Interconnect (QPI) technologies, we believe we can continue to develop a common architecture to address both Xeon and Itanium processors.

The Itanium roadmap appears interesting and compelling for many years to come. And many SGI customers have Itanium systems. SGI will continue to work with our current and new customers to determine the best choice for their microprocessor needs, as we continue to leverage the best world-class micro-processing architecture available.

Read my FLOPS. SGI is 100% committed to Itanium.

Note: SPEC and SPECfp are registered trademarks of the Standard Performance Evaluation Corporation. Competitive results from www.spec.org as of 7/24/2009.

Performance Matters

We are entering a new era of high performance computing which was one of the reasons we brought together Rackable Systems and Silicon Graphics to launch the new SGI era. I believe this SGI era will be marked by some important system characteristics: (a) open software (b) standardized interfaces (c) high efficiency (d) high density and (e) power optimization.

High performance is at the core of the SGI mission. By high performance I mean designing systems for the extreme: scale, performance, power and density.

We have solid proof points as we enter this new era: recent examples of a #1 open system, #1 commercial system, and #1 efficient system all running on SGI. Let us dig into this a bit.

For the last 15 plus years, www.Top500.org has assembled and maintained a list of the 500 most powerful computer systems in the world. The list ranks them by their performance on the LINPACK Benchmark. Top500 is a fantastic site authored by fantastic people.

The most recent list was published two weeks ago at ISC in Hamburg.

I would like to highlight a few systems on the list that represent the best of today's technology and reflect some of the important system characteristics mentioned above:

The #1 "open system" is Pleiades from NASA. It runs open Linux SLES10 with standard x86 Intel Xeon QC 3.0/2.66 GHz processors on SGI Altix ICE.

The #1 "commercial system" is Total Exploration Production. It runs open Linux SLES10 with standard x86 Intel Xeon QC 3.0 GHz processors on SGI Altix ICE.

The #1 "LINPACK" efficient system" is an Altix ICE 8200 from the US Army Research Laboratory (ARL).

Vector architecture dominated the efficiency metric in the past, until last year, when SGI Altix 4700 with Itanium 2 was the first scalar system to beat the vector systems for efficiency. In the most recent Top500 list  SGI Altix ICE with x86 Intel Xeon quad-core processors took the leadership position in the metric as the most efficient performance system, achieving an efficiency rate of 93.9%.

At the end of the day, performance matters!

The Shanghai Express

UPDATE:

AMD this week announced a new program - allowing customers a cost-effective way to upgrade to Shanghai (check out the video). This program, called the Opteron Upgrade Program, allows AMD users to upgrade processor technology in many existing AMD-based servers. For example, if you have a Rackable server running on "Rev F" processors, you can quickly and easily upgrade to a "Shanghai" processor. As you can see in the video, it can take fifteen minutes per server and will significantly up the performance and efficiency of your data center. Turn off your machine and replace CPUs. You will also need a BIOS upgrade.

The program includes 2,000 series and 8,000 series AMD Opteron processors, dating back to dual-core Rev F and including the widely available 45nm quad-core AMD Opteron processor and upcoming six-core Istanbul processor. I am eager to blog on Istanbul and six-core technology.

Below is a recent article I posted on Shanghai describing some of the benefits of these new processors. The upgrade program is another avenue to gain many of those benefits, but doing so, in a very cost effective way.

-------------------------------------------------------

The AMD ShanghaiHE has arrived and it is impressive (Shanghai is the code name for the third-generation AMD Opteron technology).  It is 45nmtechnology, supports 4-core processing (Quad Core), operates at high-efficiency (HE) and low-power (55 watts) while packing a punch with 6MB of L3 cache (level 3 cache is the extra cache between the CPU and main memory).  Key features include AMD-V(V for Virtualization), HyperTransport, and support for higher speed DDR2 memory (up to 800 MHz).  And it looks great.

Shanghai is important to the industry for what it represents today and for the future. 

As for today, the technology is available and proven and operates within existing socket technology (Socket F).  This allows new systems to fully leverage the feature set and investment protection for existing systems (upgrade the CPU, Memory and BIOS with known thermals and operating characteristics).

As for the future, Shanghai is also an architecture.  It is intended to serve as the foundation for the future generations of AMD processors: Istanbul, Magny-Cours, and San Paolo.  In case you are interested in the naming convention, AMD uses Formula One racing cities as the code names for their processors.

Istanbulwill be based on Shanghai and sport a single die six-core solution.  Yes, from four cores to six cores.  Say that twice really fast, and you get a dual-die 12 core (code named Magny-Cours).  Wheelock Latin scholars would label this a Duodecim-Core.  These 12 core processors will use AMD's G34 socket technology (code named Maranello). AMD also plans an 8-core solution named Sao Paulo.

A duodecim-core enabled MicroSliceCloudRack would house over 3,100 Shanghai-like compute cores.  That is a 4.8x increase in density over today's solution.  Said differently, if you have 100 cabinets in your data center today, you could shrink that down to 21 cabinets.

Rackable is proud to partner with AMD in supporting these achievements and our product linefully supports the Opteron HE series of technologies. 

Long live the Shanghai Express.

The Nehalem Cometh

Intel this week announced the arrival of their next generation Xeon processors, code named "Nehalem" and officially called the Intel Xeon Processor 5500 Series. It is a tremendous step forward with a new processor architecture, new platform architecture, new memory subsystem, new I/O system, and more SSD options (see my earlier blog: "Some Like it Hot"). Much has been written over the last few days on Nehalem, but not much has been written on the business impact and advancement over the last four years, and what Nehalem truly represents.

         

Let me start with a few business observations about the last four years.  What type of advancements has Intel delivered over that time period? There are at least three bright line advancements: consolidation, performance, and core intelligence. Let me elaborate a bit:

Consolidation: Intel has advanced single core Xeon to quad-core Xeon by delivering 9x on performance, 9:1 on server consolidation ratios, or 90% operating cost reductions, to lower the payback to only eight months. Refreshing to the 5500 series is clearly compelling.

Performance: In 2005, single core Xeon executed 51K business operations per second. With the 5500 Series, this has expanded to 447K business operations per second. All while lowering power consumption by about 10x. Note: business operations as demonstrated via SPECjbb2005 bops.

Intelligence: The intelligent power technologies have advanced about 5x over all. Operating states (frequency) has expanded from 3 to up to 15, idle CPU power has been reduced from 50W to 10W, and faster transmissions from 10 u-sec to < 2 u-sec. Users should really pay attention to this micro-level of intelligence.

Virtualization: Key advancements are two fold, better performance for more virtual machines and lower overhead for virtualization software efficiency. Three key technologies contribute to these advancements: support for higher memory speeds, the ability to store extended page tables (EPT) in the processor (not the software) and better addressing for virtual processor IDs (VPID) reducing the frequency of invalidation of TLB entries. The processor just runs faster and doesn't slow down for correction.

The summary is intelligent performance, automated energy efficiency, and flexible virtualization.

Rackable has been integrating compute, storage, networking and power distribution technologies at the cabinet level for years, with factory services loading customer operating systems and deployment images - we just don't talk about unified computing, we live it. We do this by hand picking the best subsystems and components in the industry like Nehalem.

Another reason for checking out Nehalem is its high performance nature across many workloads: applications that are bandwidth intensive, applications that are threaded, and those applications that have great utilization elasticity (very low lows and very high highs, that are not predictable). Volatility is a new thing ;)

The 5500 Series also supports a new on board chip set enabling DDR3 memory. DDR3 transfers data at twice the rate of DDR2, thus enabling higher bus rates, and at much lower power profiles. It runs 2x faster on less power. Don't forget, DDR3 works in multiples of 3, so, you should always deploy 3 DIMMs at a time.

For those interested in history, Nehalem is named after the river in Northwest Oregon. The river is named after an American Indian nation.

Lastly, though Nehalem was in the news this week, the architecture was actually launched at the end of 2008 with Intel Core i7, a desktop processor (advancing Intel Core 2).

The Nehalem Cometh.

Some Like it Hot

There are a variety of large unresolved problems in data centers today.  One worth solving is what I call the second watt problem.  For every watt it takes to power a server, there is a second watt (or more) used to cool the data center. The ratio of that second watt to the first watt is called a Power Usage Effectiveness ("PUE") ratio. Rackable helped create this metric as a founding member of the Green Grid. PUE equals Total Facility Power divided by IT Equipment Power. Consider: if you have 2.5 watts at your utility meter and only one watt is delivered to your server, that would be a PUE of 2.5, and a not so cool ratio. For example, our ICE Cube^TM modular data center solution design point is to deliver a PUE approaching 1.0, in addition to all the other benefits of modularity, mobility, and facility cost reduction.

        

We introduced CloudRack^TM C2 and Power XE^TM to our customers last week - the solutions are immediately available (see above pictures). In traditional data centers, to enable a PUE approaching 1.0, you need to: (a) reduce the power needed to cool the data center and (b) increase the efficiency of the actual server or storage node. CloudRack C2 was designed to reduce data center cooling needs and Power XE was designed to raise efficiency. Let me expand as to how these technologies work together.

CloudRack C2 standardizes on 12V distribution within our cabinet (now a standard cabinet width and height) and thus eliminates the need for power supplies. We also eliminated all fans from our trays and chassis. Power distribution is redundant and centralized at the back of the cabinet using our new Power XE technology, and heat exchange is managed by self-regulated fan arrays at the back of the cabinet as well. In our view, the benefits are as dramatic as when we introduced half-depth, back-to-back server technology in 1999. What is the bottom line?

The end result is a dramatic change in the data center set point. Most data centers run between 65 and 68 degrees Fahrenheit. The degree Fahrenheit of input air is called the set point.

With CloudRack C2 and Power XE, our server and storage technologies can run at very high set points. In fact, the thermal management and heat exchange elasticity is so great, we see set points in the high 80's to low 90's in degrees Fahrenheit - good for carbon-based life (humans) and silicon-based life (hardware). In fact, we will warrant and support operating temperatures all the way to 104 degrees Fahrenheit (40 degrees Centigrade) - Bikram Yoga for servers?

This approach will allow many data centers to be cooled with ambient outside air using air-side economizers, reducing or even eliminating the need for chiller-based cooling all together. Pretty amazing - just open the window and let the fresh air in. This could cut in half your data center power costs without changing any infrastructure.

Next, industry-standard power supply units ("PSU") deliver anywhere from 70 percent efficiency all the way up to the low 90's. Power redundancy as found in most redundant AC power supplies is normally accompanied with a significant efficiency penalty and by a significantly higher acquisition cost. In CloudRack C2, our cabinet-level conversion to 12V DC delivers industry-leading efficiencies while providing redundancy at a cost competitive price point.

Once converted to 12V DC, our Power XE distribution technology delivers this 12V power at 99 percent efficiency to all servers in the cabinet. The elimination of all power supplies at the server level makes this possible. Furthermore, this high efficiency is achieved whether the server is idle or at full utilization. The other big benefit of Power XE is that it perfectly balances phase power into the cabinet, allowing data center operators to capture back up to 20 percent of their already paid-for power. This "stranded power" concept is rather involved, and I plan to dedicate a future blog post to it.

 

I would also like to highlight, now that fans and power supply units are removed from trays and chassis (see above), more usable tray-level design space is opened up. CloudRack C2 is thus our most optimized platform for Build-to-Order, to date.

CloudRack C2 and Power XE when used together will advance the goals of a PUE approaching 1.0 in traditional data centers. One way to express this is: "Some Like It Hot."

A Solid State of the Union

Solid-state drives ("SSDs") have been in the market for some time, first leveraged in consumer devices, and then laptops, notebooks, and desktops.  SSDs are not new; in fact, they have a long 30 year history.  As capacity, reliability, and performance have improved over the years, coupled with DRAM advancements and price declines, solid-state drives have now hit a cross over point, making them available and affordable for enterprise-class server and storage solutions.  Let's explore this a bit.

What is a solid-state drive? An SSD uses either flash non-volatile memory - NAND or NOR (NAND flash has faster write and erase times than NOR), or DRAM volatile memory to emulate a traditional spindle / platter hard drive device ("HHD").  SSDs can directly replace HDDs.

Today's SSD capacities range from 32GB, 64GB, 80GB, 100GB and most recently, 160GB. In the HDD world, one measure of advancement was platter technology; replace "platters" with "planes" in the SSD world.  As planes and channels are interleaved with front and back-end caches - all colliding within a 2.5 inch-wide space - capacities will rapidly increase to 200GB, 300GB, 400GB, and 500GB.  It is my opinion that innovation dollars will rapidly transition from spindle technology to plane technology, and, as a result of that transition, the pace of SSD capacity expansion will galactically outpace HDD.  Consider, in the late 1980's, there were 5.25 inch 110MB drives.  Twenty years later, there are 2.5 inch 1TB drives (check out this great picture). It could take SSDs five years, not twenty, to make similar progress. For example, by 2011, we could see the first 1TB SSD.

What are the advantages of SSDs over HDDs? There are many and they are compelling.

First and foremost, there are no moving parts, and thus, no rotational vibration, which dramatically increases reliability.

Second, SSDs require very little power at idle (less than a watt) and maximum usage of 2 - 4 watts. In fact, the power profile is so compelling, suppliers are measuring work done using Joules (1 watt hour = 3600 J). Said differently, at idle, SSDs consume less than 108 Joules (0.3 Watts/hr).

Third, acoustically, SSDs are essentially silent.  The Threshold of Hearing ("TOH") is 0 dB. A whisper is 20 dB.  Normal conversation is 60 dB. SSDs operate between 0 and 2 dB.

Fourth, given that we are now in the world of solid state, you can run your data center at a higher temperature. I have always believed that data centers just need to turn up the thermostat, reducing the power consumption of Computer Room Air Conditioning (CRAC) units. SSDs are yet another reason to do so.

And finally, and I think the most compelling advantage: performance.  SSDs will have faster read and write speeds with near zero access times. For example, when benchmarking Intel's X25-M SSD, random read IOPS is 9x that of 7200 RPM HDDs and 11x faster for random writes (both on 4k transfer sizes). Further, sequential read performance is over 200MB/s, and sequential writes are over 70MB/s. This is pretty impressive performance, whether it is random reads, random writes, sequential reads, or sequential writes.

In sum, SSDs are smaller, cooler, faster and more reliable and now highly applicable to enterprise workloads.

To be balanced, there are some potential drawbacks, and I remember them by using the 3 C's: Cost, Capacity and Cycles.  First, there is the cost. SSDs are between 5x and 10x the cost per GB of HDDs. A year ago, the ratio was as high as 30x. Second, HDD technology is 10x the capacity of the largest capacity SSD. Third, there is a lifetime limit on the number of erase/program cycles per "cell." The component NAND cycle capability combined with an SSD's NAND management  algorithms determines the life span, or endurance of the SSD.  The quality of those NAND management algorithms varies from one SSD vendor to another.  Having said all that, cost is declining, capacities increasing, and memory write algorithms improving. To mitigate the drawbacks, just keep in mind a few simple items when considering an SSD: know your application, know the usage model (reads and writes), and know what is important (sequential or random) to optimize performance.

Lets look at some real products from Intel. Intel is currently offering three SSD solutions: the X25-M, X18-M, and X25-E.  The X25-M is targeted for mobile devices. I like the X25-M (up to 160GB MLC) for read intensive applications - fast boot drive, fast search and small but fast video bursts. The X18-M (1.8 inch form factor) is targeted at even smaller form factors. Finally, the X25-E targets enterprise solutions. The X25-E is a great SSD for write intensive, highly random workloads. It is a 2.5 inch form factor drive that delivers up to 64GB SLC (single-level cells) capacity and 240MB per second reads and 170MB per second writes. It also supports a 3.0GB/s SATA interface and has 2M hours MTBF. Intel has delivered a very impressive solution set as SSDs enter the data center. Rackable is pleased to support Intel's complete SSD line.

All things die and are replaced by another form. So it will be with HDD and SDD. We will see it first in laptops (the small and the weak go first). Next will be boot and read servers (the really fast), followed by the mainstream server (somewhere in the middle). The march to the solid-state server (no moving parts) is unstoppable.

This will indeed be a solid state of the union.