7.14.2014

Profile of an IBM Scientist: Giorgio Signorello

Who: Giorgio Signorello
Location: IBM Research - Zurich
Nationality: Italian

Focus: Materials Integration and Nanoscale Devices

“Some kids grow up and want to be a fireman or police officer - When I was kid, I wanted to be a scientist. Later on I wanted to become an architect or a designer - I loved the creative side of both - but my parents suggested that I should be a chemist or physicist. They thought it wasn't normal for someone so young to be so strong in these areas. Years later I started working in a clean room in Sweden and for the first time I felt that I made the right choice. It may sound crazy, but I feel at home in a laboratory.”

“What fascinates me the most about science is that it is not based on any one individual — it’s a collective contribution. What I discovered in my most recent Nature Communications paper will hopefully lead to another discovery years from now. In that sense, science is a giant puzzle without fixed borders and each time we publish we add one piece to the picture. 150 years ago James Clerk Maxwell described for the first time the physics of light and today we are manipulating light to send data at the nanoscale. Thousands of publications had to happen to reach this point. It’s inspiring, crazy and not far from being magical.”

Insider Tip:

“I applied to the IBM Research lab in Zurich five times over two years. Finally, my persistence paid off after emailing Heike Riel, who is now an IBM Fellow. But this taught me a lesson, which is applicable to any aspiring scientist — be persistent. Whether you are having challenges with a difficult experiment or dealing with a referee who keeps refusing your paper, don’t give up and push yourself.”

Last week Giorgio was awarded a 5000 CHF ($5600 USD) prize from the Swiss Physical Society for his his outstanding scientific achievements investigating the effect of strain on semiconducting nanowires, which can make it possible to transfer data between chips at the speed of light.

Read
Giorgio's publications and connect with him on LinkedIn.

7.07.2014

40+ Year Old Challenge Solved for Phase Change Materials


Phase change materials, were first considered for storing data in the 1970s, where the two metastable states or phases of these materials, are used to store data in the form of millions of lines of binary code made of up billions of 0s and 1s. 

The concept eventually reached the consumer market, and today the most common use of these materials is in optical storage, where the phase transition is induced by heating the material with a laser beam - this is how a Blue-ray disk stores a video.

The cross-sectional tunneling
electron microscopy (TEM) image of
a mushroom-type PCM cell
is shown in this photo.
In addition to a laser, it is also possible to heat the phase change material through electrical means by placing it between two electrically conducting electrodes. This forms the basis for a novel concept called phase-change memory (PCM), a nonvolatile memory technology that promises to bridge the performance gap between the main memory and storage electronics, spaning from mobile phones to cloud data centers.


The nanometric volume of phase change material in the PCM cell can be reversible switched from the amorphous phase (logic “0”) and the crystalline phase (logic “1”) by the application of suitable voltage pulses. The resulting data can be read out by applying a much lower read voltage.

But for more than 40 years scientists have never measured the temperature dependence of crystal growth, due to the difficulties associated with the measurements which are taken at both a nanometer length and a nanosecond time scale. That was until earlier this year when, for the first time, IBM scientists in Zurich were able to take the measurements, which is today being reported in the peer-review journal Nature Communications.


On the eve of the publication of this important result, the authors answered a few questions from their lab in the Binnig and Rohrer Nanotechnology Center at IBM.

IBM scientists Abu Sebastian, Manuel Le Gallo and Daniel Krebs

Let’s start with the obvious decades old question, what is the temperature corresponding to maximum crystal growth?

Daniel Krebs: The optimum crystal growth temperature is 477 degrees Celcius (750 Kelvin), but that it really just one point on the chart (figure B) – holistically it gets much more interesting. 

What is more useful to scientists studying phase change materials is that we were able to model the entire growth velocity curve in addition to this maximum. Prior to this paper, scientists knew some of the points, but not across such a wide temperature and time scale.

It is also worth noting that we took these measurements within the cell. Typically, experiments took place outside the cell, which then had to be extrapolated. Now scientists have an excellent reference point.

Can you describe the eureka moment?

Abu Sebastian: Let me start by saying that these phase change materials are very fascinating and possess unconventional crystallization kinetics. Just by changing the temperature by a few hundred degrees, you change the crystal growth rate by 17 orders of magnitude (that is beyond a trillion). This is why it  has been so difficult to probe experimentally.

Only in the last 18-24 months have scientists begun to probe the crystallization rate within a reasonable temperature range, until this point the measurements were at very low temperatures (close to room temperature).

Our key insight was in exploiting the nanoscale dimensions and the fast thermal dynamics of the phase change memory cell to expand the temperature range all the way up to the point at which the material melts – more than 600 degrees Celsius.


Daniel: It’s called the time-temperature dilemma. At room temperature you want stability of the material to retain the data for at least 10 years, but when you want to write to the material it needs to crystallize in nanoseconds. And that is what makes this material so interesting, but it’s also what makes it challenging – particularly in how it can be accurately measured.

Manuel Le Gallo: I came to IBM to do my Masters thesis work on electrical transport in phase change materials. One of the requirements was to achieve the same amorphous volume at all temperatures. This involved a deeper understanding of melting and crystallization in the PCM cells. As we delved more into the subject, the focus of the thesis gradually shifted, culminating in the fascinating results we present in the paper.

What inherent challenges in phase change memory does this achievement address and what are the potential applications?

Daniel: If we break down the challenges of PCM into read and write operations, in this work, we are addressing the write operation. Our measurements will help devise ways to write data faster and with better retention.  


Abu: In the context of PCM, this research will help us in estimating how fast we can write, how much power is required and what the real retention time is. Going beyond memory, yet another emerging application of phase change materials is in neuromorphic engineering,  creating chips based on the biological architectures of the nervous system. So understanding the phase change mechanism is critically important for a number of applications.

Manuel: Crystal growth and subsequent change in electrical conductance has the potential to emulate the biophysics of neurons and synapses. This will also form part of my doctoral thesis work which I am currently pursuing jointly with the Institute of Neuroinformatics at ETH Zurich.

What will you study next?

Abu: It will be interesting to look at different materials and compare the temperature dependence of crystal growth. We also discovered that the crystal growth rate reduces over time, which we want to expand on further.

Daniel: The reduction in growth over time is actually very interesting for me. In the amorphous phase the materials are a glass. Like a glass window becomes thicker when it is at rest over a long period of time, like 100 years, also our amorphous material will change. In fact, it changes in such a way that it becomes more viscous. This viscosity is one of the characteristics which determines how fast the material can crystalize. Therefore it effects the write operation. It cannot crystallize as fast anymore, which is a good thing for data retention. On the other hand the glassy nature also causes the inherent problem of resistance drift in phase change memory.


Creating secure test data to test systems


Tamer Salman
Editor’s note: This article is authored by Tamer Salman, senior researcher of Security & Quality Technologies at IBM Research – Haifa.

How does an insurance company, financial institution, healthcare organization, or government body get personal and confidential data to develop and test the performance of new software? The challenges associated with managing personal and confidential data are huge, especially with the increasingly stringent data privacy regulations. Some data is private and confidential, other data may have been redesigned and transformed, and some may not exist at all. Typically, project leaders or database administrators will set up an additional environment for development and testing. The big challenge is how to populate it with data.

With expertise in constraint satisfaction and automatic test generation, IBM researchers in Haifa developed the Data Fabrication Platform (DFP). It’s a solution that efficiently creates high quality test data, while eliminating any potential data security or privacy concerns. The platform is already helping a large insurance company revamp their current processes around test data.

Generating masses of personal (but fabricated) data

For most situations, generating the mass of data needed involves in-house scripting, simple data generation techniques, manual insertions and updates, and a lot of masking and data scrubbing. Even after the test data is ready, the requirements can change during development, rendering the current data useless and necessitating a repeat of some processes. The result is a tedious, costly, and time consuming process that doesn’t necessarily deliver results.

In order to accommodate distributed and outsourced development and testing, our client needed test data that would not be susceptible to leaks or breaches in security and privacy. They also need the ability to transform and evolve the data as the business needs changed or were updated. DFP does this by allowing for rule sharing and migration. It also minimizes test-data generation efforts by eliminating security and privacy concerns, and offering support for early development and regression tests. 



Data rules

The logic of what’s needed in these secure, confidential instances can be described using various rules that define the relationships between different columns in your databases, resources for populating new data columns, or transformations from archived data. DFP lets companies put these rules into the system, and get the data needed as output. The platform consumes the provided rules and generates the requested data, which can be automatically inserted into the target databases, or any of a variety of formats, such as XML, CSV, and DML files.

At the heart of the DFP lies a powerful Constraint Satisfaction Problem (CSP) solver, also developed in Haifa. A CSP typically involves numerous possibilities that can’t be solved automatically by straightforward algorithms within an acceptable amount of time. A form of artificial intelligence, the CSP solver from IBM solves these unique complex problems using it's ability to arrive at many more buildable solutions than traditional optimization approaches. The CSP solver provides accelerated solutions and helps eliminate errors by generating only data that is valid for the specific requirements.

In summary , the IBM Data Fabrication Platform is an easy to use technology that allows for rule sharing and migration, minimizes test-data generation efforts, eliminates security and privacy concerns, and makes it easier for companies to outsource development and testing. 

6.26.2014

SERI Predicts Improved Utilities of the Future


Data and analytics hold the key for many global industries facing a future with difficult obstacles and new opportunities. Few industries are facing challenges as significant as those in the energy and utilities field.

Today’s aging energy infrastructure isn’t just being battered by dramatic weather events. Increases in cyber attacks on the grid have skyrocketed – from 2012 to 2013, more than 53 percent of cyber attacks have been on energy installations.

Dr. Jonathan Pershing
“Energy demand is rising, but supply is dropping … and our aging infrastructure is vulnerable,” said Dr. Jonathon Pershing, the principal deputy director for the Office of Energy Policy and Systems Analysis at the US Department of Energy, who keynoted the second annual Smarter Energy Research Institute Conference at IBM Research, before more than 125 energy and utility experts from companies around the world

SERI was formed in 2012 with the goal of bringing together the world’s top energy and utility companies to build the energy utility of the future using data analytics. It pairs IBM’s open analytics toolkit platform of application-specific code with energy and utility companies’ ideas, needs and expertise to develop new software applications that solve their operational problems. At this year’s conference, Institute partners Alliander, Hydro-Québec, DTE Energy and IBM spent two days with representatives from 18 companies from around the world to learn how the big data analytics applications demonstrated at the conference can address these challenges.

In 2013, the US averaged 140 minutes per electrical outage, its highest rate in history.

In his keynote Pershing also pointed out that “50 percent of natural gas transmission infrastructure in the US was built between the 1940s-60s … and by 2030 the US will need to invest between $1.5 and $2 trillion in utility improvements.”

He said this comes at a time when more than 60 percent of those working in this industry are likely to retire in the next decade. As part of Pershing’s role with the DOE, he is leading their Quadrennial Energy Review Project, a four year effort to improve the country’s energy production, generation, supply and demand, and entire value chain. And he’ll count on the innovation of groups such as SERI to deliver solutions to these needs.

Predictive Analytics: Energy solutions through smart software

Every country faces similar energy challenges. DTE Energy, Alliander, and Hydro-Québec are using software – big data analytics – to better understand and solve these challenges.

Take weather prediction. Data, and lots of it, is behind the collaborative project Outage Prediction Response Optimization (OPRO) application, led by SERI partner DTE Energy. It is designed to predict weather and potential damage to the grid in order to help the utility optimize resources and prepare its response proactively. OPRO can predict where the storm will hit, down to one kilometer; who it will affect; and even utility service response times.

OPRO screenshot
The weather isn’t the only thing utilities need to predict. Grid maintenance costs in the US increased 42 percent between 2011 and 2012. Another DTE Energy-SERI project, Asset Risk Management and Optimized repair-rehab-replace (Armor^3), applies predictive and prescriptive analytics to grid data to quantify and optimize infrastructure maintenances planning for all electrical assets, including transformers, cables, poles, circuits, and more.

SERI partners are working on a handful of other applications that, among other things, use data to optimize the management and preservation of energy we depend on from all sources. Energy and utility companies interested in joining SERI can read more about how it’s building the utility of the future, and about some of the partners’ projects already underway.

This article is by Cody Frankel, a senior at the University of Rhode Island's Harrington School and intern with IBM Research Marketing and Communications.

6.25.2014

IBM and RWTH Aachen University Collaborate on Phase Change Materials with Support from European Commission


Kaes and Krebs working in the
Binnig and Rohrer Nanotech Center
at IBM Research (left to right)
Several years ago the European Commission created the Marie Skłodowska-Curie Actions (MSCA). It set aside €6.16 billion to last through 2020 to support the research training and career development of 25,000 European researchers whether they are early PhD candidates or seasoned scientists. The goal is to encourage trans-national and interdisciplinary mobility across member and partner countries, and for scientists to exchange ideas which lead to innovation.

One of these candidates is Matthias Kaes, a pre-doctorial student from RWTH Aachen University in Western Germany. For the past eight months, Matthias has been working side-by-side with scientists at IBM Research-Zurich on the cleverly titled project called DIASPORA or Drift In Amorphous Semiconductors – A Partnership of Rüschlikon and Aachen (Rüschlikon is the Zurich suburb where IBM is located). The collaborative project is focused on the emerging field of phase change materials, a promising class of materials which are expected to be the basis for new generation of fast and durable memory to handle Big Data and cloud computing workflows.

As Matthias’ time at IBM winds down to make room for other students joining from Aachen, we had a chance to ask him, and the lead scientist on the project Dr. Daniel Krebs about the experience.

What motivated you to apply for MSCA?

Daniel: As an alumnus of Aachen, we have very good connections to the phase change materials team who are doing some impressive and complementary work to our research. But we didn’t collaborate on a concrete project, so the MSCA call gave us the chance to formalize a project, with dedicated resources, to investigate some critical basic science which has been vexing phase change materials scientists since the 1970s. I was personally also motivated by the opportunity to interact with young minds and to exchange ideas, which is the goal of the MSCA.

What did you know about IBM Research prior to joining the team here? Any surprises?

Matthias: Of course I knew about the Nobel Prize research and the invention of the scanning tunneling microscope, which led to the atomic force microscope. But I was very surprised about the other research conducted here, including computer security. I was also surprised to learn about the large number of PhD students, I didn’t think this was possible at an industrial lab.

Tell us more about your experience so far?

Matthias: A typical day is spent in the lab where I am setting up a variety of experiments, testing, calibrating and characterizing the phase change materials under different conditions. I also participate in weekly team meetings with our co-workers in Aachen.

One of the goals of the MSCA is your personal development. Do you have any examples?

Matthias: It may come across as strange, but I have noticed that compared to my time at the university I am taking a lot more notes. We have weekly cadence calls with the other students, both here and in Aachen, and for these calls I need to update the team on my progress. Therefore, since I am presenting more than before, I obviously need to prepare with shorter deadlines. I find this a good "best practice" for my career. 

In addition, the research here is more focused. For example, I am working with one experimental set-up, compared to several in Aachen. This is enabling me to find a niche to focus on in more detail.


Based on your experience, how do you think about a career in academic or industrial research?

Matthias: I had the feeling before I arrived that I didn’t want to stay in academia, so this experience has reinforced my initial thinking. My experience has also opened my eyes to the level of pressure for results at an industrial lab. It’s too early to know where I will go, but I would like to at least initially start at an industrial lab once I finish my PhD.

Since you are the first of several students from Aachen joining IBM what tips can you offer?

Matthias: Well, you should be prepared to have patience finding a flat demand far exceeds supply! But one thing that has really helped me to meet people and to feel comfortable in my temporary home was to join the IBM football team, which is sponsored by the lab’s Hobby Club. It was good to meet colleagues outside of work and to meet scientists from the other IBM departments. So my best advice would be to take advantage of this.

6.24.2014

Dealing with errors in quantum computing


Jerry Chow
Editor’s note: This article is by Jerry Chow, manager of Experimental Quantum Computing at IBM Research.

Quantum computers promise to open up new capabilities in the fields of decryption and simulation not possible in today’s computers. And when made a reality, the performance improvement will be due to their fundamental unit of information: a qubit. Qubits are two-level systems that obey the laws of quantum mechanics. Imagine emulating a quantum computer using today’s approaches to building the largest computer systems in the world. If a quantum computer could be built with just 50 qubits, there is not a combination of today’s Top500 supercomputers that could successfully emulate it. The holy grail of quantum applications is to perform tasks like large number factorization and simulation of complex quantum systems, problems which are intractable with today’s supercomputers.

But like today’s machines, quantum computers suffer from errors, and, worse, these errors seem to be fundamental, since quantum information is so fragile. Our team at the Thomas J Watson Research Center published results in the paper Implementing a strand of a scalable fault-tolerant quantum computing fabric (doi: 10.1038/ncomms5015) in Nature Communications (1) about recent experimental steps toward a “surface code” that shows promise for correcting these errors – and bringing fault-tolerant quantum computers a step closer to reality.

Understanding a qubit’s peculiar properties

3-qubit, 5 resonator device
The classical equivalent of a qubit is the digital bit, the “1” and “0” ubiquitous in all modern computers. Qubits, though, can exist in some combination of 0 and 1 simultaneously, a different state altogether that is called a “superposition.” When qubits interact with each other, they can form a special kind of superposition that is called entanglement. Entangled states exhibit perfect correlation no matter how far the qubits are separated in space, and this may be one of the phenomena that grant quantum computing its power.

Entanglement is necessary for quantum computing, but can also lead to errors when it occurs between the quantum computer and the environment (i.e. anything that is not the computer itself). Quantum effects disappear when the system entangles too strongly to the external world, which makes quantum states very fragile. Yet, there is a kind of tension, since the quantum computer must be coupled to the external world so the user can run programs on it and read the output from those programs.

This need to couple the quantum computer to its environment sets a limit to how well the system can maintain its quantum behavior. And as it interacts more strongly with the world, errors are introduced in the computation. How long a qubit retains its quantum properties is referred to as the coherence time and is a common metric to benchmark the quality of a qubit. The art therefore lies in building quantum systems with reduced errors and long coherence times.

Quantum error correction theories

In order to build a fully-functional, large-scale, universal, fault-tolerant quantum computer, we will need to figure out how to have long coherence times and deal with errors which may arise from the manipulation of the quantum computer. The path forward is via quantum error correction (QEC), a robust theory which has been developed from the ideas of classical error correction in order to deal with errors in qubits. In classical error correction, a bit (taking values 0 or 1) is encoded into multiple physical bits. For example, three physical bits, 000, can encode the logical bit value of 0. If any one of the physical bits happens to flip its state because an error has occurred (001), the original logical value (0) can still be recovered by “majority voting” (two “00s” overrule the “1”). 

Encoding qubits are substantially more challenging than a bit. For one, they can’t be cloned, so we cannot simply copy to an analogous “000” bit state. We also can’t “see” the quantum information in the same way because looking at or measuring the qubit, which could be in any superposition of “0” or “1”, forces the state to choose either “0” or “1”. But it turns out that all of these problems can be overcome by the clever use of entanglement and superposition.

A magnified look at a single quantum bit

QEC protocols rely on parity measurements. An example of a QEC protocol that protects a logical qubit from a single bit-flip error is the three-qubit Shor code. In that code, via superposition and entanglement, pairs of qubits in a three qubit register can be interacted with in such a way to give parity information, (i.e. are both qubits either 00 or 11 having even parity, or are both qubits either 01 or 10, having odd parity). Through the accumulation of this parity information from the register, it is then possible to detect and locate a single error in the qubit register.

But to make a fully-fledged quantum computer work, we need codes that protect against a continuum of errors on multiple qubits. Our team is focused on surface code. It has a high error threshold and only nearest-neighbor parity checks. This means that error rates do not need to be excessively low to see the benefits of coding, and each operation we need to do only involves a few physically adjacent qubits. This makes the surface code an attractive option for an experimental demonstration with superconducting qubits.

Surface code with superconducting qubits

We have been exploring superconducting qubits to build a universal quantum computer based on the surface code architecture for quantum error correction. Because their properties can be designed and manufactured using standard silicon fabrication techniques, we anticipate that once a handful of superconducting qubits can be manufactured reliably and repeatedly, and controlled with low error rates, there will be no fundamental obstacle to scaling up to thousands of qubits and beyond.

Coherence times for superconducting qubits have been increasing steadily for the past 10-15 years, and in 2010, those values, together with the ability to couple and control multiple qubits with low error rates, reached a point where we could start to consider potentially scalable architectures. In our newest paper, we combined a number of state-of-the-art advances within superconducting qubits in order to demonstrate a crucial stepping stone towards the surface code quantum error correction architecture. Using a three superconducting qubit network, we successfully detected the parity of two “code” qubits via the measurement of a third “syndrome” qubit (the “error detection” qubit). A larger surface code system would involve similar parity checks as we have demonstrated in this reduced system.

Our result and recent findings on high accuracy controls from UC Santa Barbara show the promise for superconducting qubits. The architectural and engineering challenges that lay ahead are ripe to be addressed to get towards a fault-tolerant quantum computer.


(1) Implementing a strand of a scalable fault-tolerant quantum computing fabric  

IBM Thomas J. Watson Research Center: Jerry M. Chow, Jay M. Gambetta, Easwar Magesan, David W. Abraham, Andrew W. Cross,
Nicholas A. Masluk, John A. Smolin, Srikanth J. Srinivasan, M. Steffen 


Raytheon BBN Technologies: B.R. Johnson, Colm A. Ryan

6.23.2014

Rethinking HPC Benchmarks with a Focus on Applications and Power Efficiency

From left to right: HPC scientists Costas Bekas, IBM;
Pavel Klavik,IBM and Charles University in Prague;
Yves Ineichen, IBM and Cristiano Malossi, IBM
Today with the release of the latest TOP500 supercomputing list we see the largest and most powerful high performance computers in the world. And while it is impressive, it’s quickly becoming an antiquated measuring stick for several reasons including how to measure workload and power efficiency.

Since 1993 the TOP500 list has been published to portray an exponential increase in performance, due to Moore’s law, but in the last 10 years, the core technology has dramatically changed with chip manufacturers turning to a recipe of multi-core chips and parallelism to keep the effect of Moore’s Law alive.

In doing so a new set of critical issues and questions have arisen about the ability to program such machines and how to deal with the very fast increase in power requirements.

As the industry moves towards more data-centric computing workloads IBM scientists believe that the answer to both questions are closely connected and have recently published a paper in the Proceedings of The Royal Society A on this topic. Some of the paper's authors answered a few questions about their motivation.

Your paper mentions the shortcomings of the FLOPs performance metric. What are they?

Costas Bekas (
@CostasBekas): The main shortcoming of the FLOPs metric is that it concentrates on doing as many computations as possible on a given piece of hardware in the unit of time. This is not about trying to solve the problem quickly, but essentially measuring how many FLOPs we can squeeze out of the machine. This metric stems from the days when computing was very expensive — so the thinking was you better use every last transistor.

But today with the cloud and massive high performance computing (HPC) machines the equation has flipped and energy is a significant cost factor. Today, the real issues are the results. Do you get the results fast and how accurate are they?

So the question clients need to ask is, do you care that you run a machine to its full "nominal" capacity or that you get the results back as fast as possible and at the lowest possible cost? The answer lies in both hardware improvements, such as water-cooled design and photonics, combined with advanced algorithms.

What should a new benchmark look like for measuring high performance computers?

Cristiano Malossi: The biggest challenge will be to find a metric which is accepted by everyone. From what I hear at various conferences is that the industry is in agreement that the LINPACK benchmark isn’t sufficient anymore. The GRAPH500 is a good start, but we need to take it a step further. I see the need to have more than one metric, in fact 13 or one for each kernel, which are then weighed, scored and averaged. This is the bottom-up approach popularized by Phil Collela's (or the Berkeley) dwarfs.

Costas: The big drivers are the HPC applications. The nature of applications has changed and there are hundreds. In the past the focus was numerical linear algebra, but today applications span bioinformatics, finance, traditional engineering, and data analytics. A new benchmark needs to account for speed to result and energy efficiency to result.

Besides providing a ranking tool for marketing purposes, do benchmarks also help the development of HPC?

Costas: Absolutely, the rankings provide scientists and engineers with a guide on designing the future systems. For instance, here in Zurich we worked at a very early stage with the prototype of the IBM BlueGene/Q. The goal was to incorporate in the design loop as soon as possible feedback from application developers and algorithm specialists. So a benchmark is more than just touting first place, it helps to close the circle between developers and users.

We shouldn’t forget that benchmarks are also important for politics and financial officials. These machines require budget after all and when using tax payer funding you need to demonstrate results and coming in the top of a benchmark is one way to do this. But a better way to demonstrate the ROI of a system is to solve a problem quickly and efficiently — designing a new drug, simulating a city to make it run more efficiently or engineering a safer product. Ask the average citizen if they prefer a blue ribbon for first place or a city with less traffic and the answer is obvious.

Your paper proposes a new benchmark based on application, time to result and energy efficiency. How can HPC users begin testing your benchmark on their systems?

Costas: In this publication and our previous paper are the starting points. Together with my colleague, IBM Fellow Alessandro Curioni (
) we have been evangelizing the need to move away from using the LINPACK benchmark alone for several years. This paper provides a framework and tools for accurate, on-chip and on-line power measurements in the context of energy-aware performance. The tools described the easy implementation of user code for the detailed power profiling. Beyond just the benchmark also describe a patented framework that enables users to combine a bulk of relaxed accuracy, non-reliable, calculations together with a small fraction of accurate (reliable) calculations in order to achieve full final accuracy. The result is up to 2 orders of reduction in energy to solution. They key is to change the computing paradigm.

The key takeaway is the following. Within 10 years we will have HPCs running billions of threads. For comparison, our world record work last year, which led to the Gordon Bell Prize (together with Alessandro and colleagues from ETH Zurich, Technical University of Munich and the Lawrence Livermore National Lab) was using 6.4 million threads. So with a billion you have tremendous scale, therefore you need to develop algorithms which can recognize failure, resolve failure, but also live with it for a reliable result.

Where does academia fit into this new benchmark? Clearly we need to teach this to the next generation of HPC users.

Costas: Indeed, and we are seeing signs of progress. For example, Pavel Klavik, one of the authors of this paper was an intern at our lab from the Charles University in Prague. He joined us after winning our annual IBM Great Minds Competition and he obviously understands where this trend is going.

Cristiano: These concepts are starting to be integrated into part of the academic curricula. In fact, we are participating in an EU project with several universities called Exa2Green with the goal of creating energy-aware sustainable computing for future exascale systems. This will help trickle down these ideas to the classroom.

Yves Ineichen: I think we are in a transition period as the topic begins to reach the lecture halls. Part of the challenge is that many universities don’t have access to these HPC systems at the level required to actually run these algorithms and testing, but with collaborations this will change. But first we need a consensus from the HPC industry.