IBM researchers announce major breakthrough in phase change memory

For years, scientists and researchers have looked for the so-called Holy Grail of memory technology — a non-volatile memory standard that’s faster than NAND flash while offering superior longevity, higher densities, and ideally, better power characteristics. One of the more promising technologies that’s been in development is phase-change memory, or PCM. IBM researchers announced a major breakthrough in PCM this week, declaring that they’ve found a way to store up to three bits of data per “cell” of memory. That’s a significant achievement, given that previous work in the field was limited to a single bit of data per memory cell.

Phase change memory exploits the properties of a metal alloy known as chalcogenide. Applying heat to the alloy changes it from an amorphous mass into a crystal lattice with significantly different properties, as shown below:


Scientists have long known that chalcogenide could exist in states between crystal lattice or amorphous, but building a solution that could exploit these in-between states to store more memory has been extremely difficult. While phase-change memory works on very different principles than NAND flash, some of the problems with scaling NAND density are conceptually similar to those faced by PCM. Storing multiple bits of data in NAND flash is difficult because the gap between the voltage levels required to read each specific bit is smaller the more bits you store. This is also why TLC NAND flash, which stores three bits of data per cell, is slower and less durable than MLC (2-bit) or SLC (single bit) NAND.

IBM researchers have discovered how to store three bits of data in a 64K array at elevated temperatures and for one million endurance cycles.

“Phase change memory is the first instantiation of a universal memory with properties of both DRAM and flash, thus answering one of the grand challenges of our industry,” said Dr. Haris Pozidis, an author of the paper and the manager of non-volatile memory research at IBM Research – Zurich. “Reaching 3 bits per cell is a significant milestone because at this density the cost of PCM will be significantly less than DRAM and closer to flash.”

Here’s how the PR blast describes the breakthrough:

To achieve multi-bit storage IBM scientists have developed two innovative enabling technologies: a set of drift-immune cell-state metrics and drift-tolerant coding and detection schemes.

More specifically, the new cell-state metrics measure a physical property of the PCM cell that remains stable over time, and are thus insensitive to drift, which affects the stability of the cell’s electrical conductivity with time. To provide additional robustness of the stored data in a cell over ambient temperature fluctuations a novel coding and detection scheme is employed. This scheme adaptively modifies the level thresholds that are used to detect the cell’s stored data so that they follow variations due to temperature change. As a result, the cell state can be read reliably over long time periods after the memory is programmed, thus offering non-volatility.

“Combined these advancements address the key challenges of multi-bit PCM, including drift, variability, temperature sensitivity and endurance cycling,” said Dr. Evangelos Eleftheriou, IBM Fellow.

There’s still a great deal of work to do before phase-change memory can be considered as a candidate to replace NAND flash or DRAM in certain situations. The performance and power impact of these new structures has not been characterized and the switching time hasn’t been revealed.


The graphic above is from an IBM video explaining how PCM memory works and some general information on this latest breakthrough. Note that PCM, like NAND flash, takes a performance hit when it shifts to a multi-bit architecture. While single-bit PCM is nearly as fast as DRAM (according to IBM), multi-bit PCM is significantly slower. Data retention (how long data remains in the cell) was also worse than NAND flash, which has lower endurance (how many read/write cycles the cells can withstand) but higher data retention.

Phase-change memory is theoretically capable of replacing DRAM in at least some instances, but if these density gains come at the cost of programming speed, the net gain may be minimal. Phase-change memory also requires large amounts of power to program and generates a great deal of heat as a result.

This video from IBM walks through the history of phase-change memory, explains the basics of its function, and covers the most-recent breakthrough. We think IBM’s discovery here could help pave the way for a long-term replacement to NAND flash, but we’re still years away from that. Intel’s Optane 3D XPoint memory may make its own play for the server and data center space, and Micron, which used to manufacture PCM for the mass market doesn’t build it anymore.

IBM Watson amps up Moogfest 2016 with AI-infused programming

Watson IBMIBM Watson came to Moogfest 2016, but there were no Jeopardy! questions this time around. If you’ve been following, you already know that IBM Watson, an artificially intelligent system capable of answering questions in natural language, has been up to much more than that recently. At Moogfest, IBM Watson team spokesperson Ally Schneider was on hand to outline all of the latest developments.

Everyone remembers Watson from its Jeopardy! performance on television in 2011. But work on the project was started much earlier — not just in 2006, when three researchers at IBM first got the idea to build a system for the game show, but really decades before that, as IBM began doing work on natural language processing and cognitive computing in the 1970s.


The Jeopardy! Watson system in 2011 had three main abilities, as Schneider explained. First, it could understand unstructured text. “[Normally] we don’t have to think about it, but we inherently understand what sentences are, and how verbs, nouns, etc. come together to produce text,” Schneider said. Watson could read through human-generated content and parse it in a way that other systems haven’t been able to do before. Next, Watson could come up with its own hypotheses, and then return the one with the highest confidence. Finally, there’s a machine learning component — one that’s not hard-coded or programmed, but that really learns as it goes. “When you were back in school, not too long ago for some, how did your teachers test you to see if you understood what you were reading?” Schneider asked. “They would give you feedback on your answers. [For example], yes, full credit… maybe you got partial credit… or no, incorrect, here’s what you should have done instead.” Watson is able to “reason” in the same manner.

Today, after continuous improvements, Watson consists of 30 open-source APIs across four categories: language, speech, vision, and data insights. “Watson [today] has the ability to read through and understand unstructured data like a human and pull out the relevant answers and insights and now images,” Schneider said. She then began to illustrate some recent examples of Watson’s power. The first and arguably most significant one was a joint effort with Memorial Sloan Kettering Cancer Center. The goal was to train Watson to think like a doctor, in order to assist oncologists working with breast and colon cancers. IBM’s team fed Watson a steady diet of medical journals, clinical trial results, encyclopedias, and textbooks to teach it the language of medicine.

From there, Watson could look at a patient’s individual information and compare it against what the system knows about medicine, and then come back with recommended treatment options. Schneider said it’s still up to the doctor to decide how to use that information; it’s not a question of man versus machine, but rather, how machines can enhance what humans can already perform. In this case, the goal was to empower doctors so that they don’t have to read an impossible 160 hours worth of material each week — an actual estimated figure for how much new research is being published on a weekly basis!

Watson Logo

Next up was an application for the music industry. Quantone delivers in-depth data on music consumption. It not only leverages structured metadata the way Pandora, Spotify, and other music services do, such as the genre of music, the number of beats in songs, and so on, but using IBM Watson technologies, it can also process unstructured data, such as album reviews, artist-curated content, and natural language classification. Using Quantone, as Schneider put it, an end user can say, “I’m looking for a playlist reminiscent of Michael Jackson from a certain time period,” and get an answer that also pulls in and considers unstructured data.

Content creators can also benefit from AI-infused programming. Sampack offers algorithmically and artistically generated samples that are royalty-free. It’s essentially an automated license-free music sample generator. It takes in descriptions of tones (such as “dark” or “mellow”) and then translates them into an audio sample using Watson’s Tone Analyzer capability. Sampack can understand descriptions and emotions and translate them into music effects, sounds, and filters.

IBM also published a cookbook recently, which as Schneider pointed out isn’t something you would have expected to hear before it happened. The book is called Cognitive Cooking with Chef Watson: Recipes for Innovation from IBM & the Institute of Culinary Education. Watson would analyze the molecular construction of foods, figured out what goes well together, take in inputs such as specific ingredients and what to exclude (such as gluten or other allergy triggers), and then create 100 new recipes using that query. It doesn’t search through an existing recipe database for these, either; instead, it creates 100 new recipes based on your inputs. The first recipe is usually pretty normal; by the time it gets to recipe 100, it’s “a little out there,” as Schneider put it.

In the art world, World of Watson was a recent exhibit (pictured below) by Stephen Holding in Brooklyn, in collaboration with IBM Watson using a deviation of a color API. Watson mined through Watson-specific brand imagery and came up with a suggested color palette for Holding to use. The goal was to invoke innovation, passion, and creativity with an original piece of art.

Stephen Holding IBM Watson World of Watson Art

Finally, IBM Watson partnered with fashion label Marchesa for the recent Metropolitan Museum of Art gala with model Karolina Kurkova. Watson was tasked with coming up with a new dress design that was “inherently Marchesa and true to the brand.” Watson was involved in every step of the way. Using another color deviation API, Watson mined through hundreds of images from Marchesa, including model photos, to get a feel for the color palette, Schneider said. Then Inno360 (an IBM Watson ecosystem partner) used several APIs and considered 40,000 options for fabric. With inputs from Marchesa that were consistent with the brand, but while also evaluating fabrics that would work with embeddded LEDs, Watson came up with 35 distinct choices. The third step involved embedding the LED technology into the dress using the tone analyzer, with specific colors being lit up through the flowers.


Today, anyone can get started working with IBM Watson by heading to IBM BlueMix and signing up for a Watson Developer Cloud account. Back in February 2015, IBM boostedWatson Developer Cloud with speech-to-text, image analysis, visual recognition, and the ability to analyze tradeoffs between different drug candidates. In July last year, Watson gained a new Tone Analyzer that could scan a piece of text and then critique the tone of your writing. We’ve also interviewed IBM’s Jerome Pesenti on many of the latest Watson developments.