Phase Change Memory Shows Promise for AI Use Says IBM

By John Russell

October 24, 2018

Research into phase change memory (PCM) technology and its application to neuromorphic computing isn’t new, but it has steadily progressed with IBM playing an important role. Earlier this month, IBM researchers published a progress report in the AIP Journal of Applied Physics describing how PCM devices could implement select AI functions effectively and at lower power than traditional CMOS von Neumann circuits.

“We are on the cusp of a revolution in AI,” argue the authors. Near-term AI will depend on traditional CMOS circuits but then shift to include non-von Neumann ‘co-processing’ approaches.

“By augmenting conventional computing systems,” they wrote, “these systems could help achieve orders of magnitude improvement in performance and efficiency. In summary, we believe that we will see two stages of innovations that take us from the near term, where the AI accelerators are built with conventional CMOS, towards a period of innovation involving the computational approaches presented in this article.”

Time will tell if they are correct. The IBM researchers take pains to show how PCM devices can be used for in-memory computing functions such as matrix-vector multiplication, unsupervised learning, deep learning, spiking neural networks, and more. They also identify and briefly discuss the challenges, however the concrete examples presented suggest PCM technology is nearing viability for use in AI and other in-memory computing functions.

The IBM paper – Brain-inspired computing using phase-change memory devices by Abu Sebastian, Manuel Le Gallo, Geoffrey W. Burr, Sangbum Kim, Matthew BrightSky, and Evangelos Eleftheriou – is an interesting accessible read.

This illustration taken from the paper (below) shows their view of how technology use is divided among AI categories.

There’s also an article on the work (A New Brain-Inspired Architecture Could Improve How Computers Handle Data and Advance AI) posted on the AIP Publishing site in which Sebastian emphasized that executing certain computational tasks in the computer’s memory would increase the system’s efficiency and save energy:

“If you look at human beings, we compute with 20 to 30 watts of power, whereas AI today is based on supercomputers which run on kilowatts or megawatts of power,” Sebastian said in the article. “In the brain, synapses are both computing and storing information. In a new architecture, going beyond von Neumann, memory has to play a more active role in computing.”

The IBM team drew on three different levels of inspiration from the brain. The first level exploits a memory device’s state dynamics to perform computational tasks in the memory itself, similar to how the brain’s memory and processing are co-located. The second level draws on the brain’s synaptic network structures as inspiration for arrays of phase change memory (PCM) devices to accelerate training for deep neural networks. Lastly, the dynamic and stochastic nature of neurons and synapses inspired the team to create a powerful computational substrate for spiking neural networks.

By way of background, phase change memory is “based on the property of certain compounds of Ge, Te, and Sb that exhibit drastically different electrical characteristics depending on their atomic arrangement. In the disordered amorphous phase, these materials have very high resistivity, while in the ordered crystalline phase, they have very low resistivity.”

As described the authors, “When a current pulse of sufficiently high amplitude is applied to the PCM device (typically referred to as the RESET pulse), a significant portion of the phase change material melts owing to Joule heating. The typical melting temperature of phase-change materials is approx. 600 C. When the pulse is stopped abruptly so that temperature inside the heated device drops rapidly, the molten material quenches into the amorphous phase due to glass transition. In the resulting RESET state, the device will be in a high resistance state if the amorphous region blocks the bottom electrode.”

PCM has two critical properties governing its use in circuits:

  • The first key property of PCM that enables brain-inspired computing is its ability to achieve not just two levels but a continuum of resistance or conductance values. This is typically achieved by creating intermediate phase configurations by the application of suitable partial RESET pulses
  • The second key property that enables brain-inspired computing is the accumulative behavior arising from the crystallization dynamics… [O]ne can induce progressive reduction in the size of the amorphous region (and hence the device resistance) by the successive application of SET pulses with the same amplitude. However, it is not possible to achieve a progressive increase in the size of the amorphous region. Hence, the curve shown in Fig. 3(c) typically referred to as the accumulation curve, is unidirectional.

Using these properties it is possible to implement a variety of logical, arithmetic, and machine learning functions with PCM memory.

Compressed sensing and recovery, say the researchers, is one of the applications that could benefit from a computational memory unit that performs matrix-vector multiplications. They note:

“The objective behind compressed sensing is to acquire a large signal at a sub-Nyquist sampling rate and subsequently reconstruct that signal accurately. Unlike most other compression schemes, sampling and compression are done simultaneously, with the signal getting compressed as it is sampled. Such techniques have widespread applications in the domains of medical imaging, security systems, and camera sensors.

“The compressed measurements can be thought of as a mapping of a signal x of length N to a measurement vector y of length M < N. If this process is linear, then it can be modeled by an M N measurement matrix M. The idea is to store this measurement matrix in the computational memory unit, with PCM devices organized in a cross-bar configuration [see Fig. 6(a)]. This allows us to perform the compression in O(1) time complexity. An approximate message passing algorithm (AMP) can be used to recover the original signal from the compressed measurements, using an iterative algorithm that involves several matrix-vector multiplications on the very same measurement matrix and its transpose. In this way, we can also use the same matrix that was coded in the computational memory unit for the reconstruction, reducing the reconstruction complexity.”

Shown here is an experimental illustration of compressed sensing recovery in the context of image compression. A 128 x128 pixel image was compressed by 50% and recovered using the measurement matrix elements encoded in a PCM array.

A key challenge with computational memory, concede the researchers, is the lack of high precision. Even though approximate solutions are sufficient for many computational tasks in the domain of AI, there are some applications that require that the solutions are obtained with arbitrarily high accuracy.

“Fortunately, many such computational tasks can be formulated as a sequence of two distinct parts. In the first part, an approximate solution is obtained; in the second part, the resulting error in the overall objective is calculated accurately. Then, based on this, the approximate solution is refined by repeating the first part. Step I typically has a high computational load, whereas Step II has a light computational load. This forms the foundation for the concept of mixed-precision in-memory computing: the use of a computational memory unit in conjunction with a high-precision von Neumann machine.23 The low-precision computational memory unit can be used to obtain an approximate solution as discussed earlier. The high-precision von Neumann machine can be used to calculate the error precisely. The bulk of the computation is still realized in computational memory, and hence we still achieve significant areal/power/ speed improvements while addressing the key challenge of imprecision associated with computational memory),” they wrote.

You get the flavor of the paper. It is an interesting overview, best read directly, showing the promising progress made in formulating ideas for structuring PCM and PCM-cum-traditional circuits for implementing AI functions.

Link to paper: https://aip.scitation.org/doi/10.1063/1.5042413

Link to AIP article on the work: https://publishing.aip.org/publishing/journal-highlights/new-brain-inspired-architecture-could-improve-how-computers-handle

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

Pittsburgh Supercomputer Powers Machine Learning Analysis of Rare East Asian Stamps

September 27, 2021

Setting aside the relatively recent rise of electronic signatures, personalized stamps have been a popular form of identification for formal documents in East Asia. These identifiers – easily forged, but culturally ubi Read more…

Purdue Researchers Peer into the ‘Fog of the Machine Learning Accelerator War’

September 27, 2021

Making sense of ML performance and benchmark data is an ongoing challenge. In light of last week’s release of the most recent MLPerf (v1.1) inference results, now is perhaps a good time to review how valuable (or not) Read more…

Quantum Monte Carlo at Exascale Could Be Key to Finding New Semiconductor Materials

September 27, 2021

Researchers are urgently trying to identify possible materials to replace silicon-based semiconductors. The processing power in modern computers continues to increase even as the size of the silicon on which components a Read more…

The Case for an Edge-Driven Future for Supercomputing

September 24, 2021

“Exascale only becomes valuable when it’s creating and using data that we care about,” said Pete Beckman, co-director of the Northwestern-Argonne Institute of Science and Engineering (NAISE), at the most recent HPC Read more…

Three Universities Team for NSF-Funded ‘ACES’ Reconfigurable Supercomputer Prototype

September 23, 2021

As Moore’s law slows, HPC developers are increasingly looking for speed gains in specialized code and specialized hardware – but this specialization, in turn, can make testing and deploying code trickier than ever. Now, researchers from Texas A&M University, the University of Illinois at Urbana... Read more…

AWS Solution Channel

Introducing AWS ParallelCluster 3

Running HPC workloads, like computational fluid dynamics (CFD), molecular dynamics, or weather forecasting typically involves a lot of moving parts. You need a hundreds or thousands of compute cores, a job scheduler for keeping them fed, a shared file system that’s tuned for throughput or IOPS (or both), loads of libraries, a fast network, and a head node to make sense of all this. Read more…

Qubit Stream: Monte Carlo Advance, Infosys Joins the Fray, D-Wave Meeting Plans, and More

September 23, 2021

It seems the stream of quantum computing reports never ceases. This week – IonQ and Goldman Sachs tackle Monte Carlo on quantum hardware, Cambridge Quantum pushes chemistry calculations forward, D-Wave prepares for its Read more…

Quantum Monte Carlo at Exascale Could Be Key to Finding New Semiconductor Materials

September 27, 2021

Researchers are urgently trying to identify possible materials to replace silicon-based semiconductors. The processing power in modern computers continues to in Read more…

The Case for an Edge-Driven Future for Supercomputing

September 24, 2021

“Exascale only becomes valuable when it’s creating and using data that we care about,” said Pete Beckman, co-director of the Northwestern-Argonne Institut Read more…

Three Universities Team for NSF-Funded ‘ACES’ Reconfigurable Supercomputer Prototype

September 23, 2021

As Moore’s law slows, HPC developers are increasingly looking for speed gains in specialized code and specialized hardware – but this specialization, in turn, can make testing and deploying code trickier than ever. Now, researchers from Texas A&M University, the University of Illinois at Urbana... Read more…

Qubit Stream: Monte Carlo Advance, Infosys Joins the Fray, D-Wave Meeting Plans, and More

September 23, 2021

It seems the stream of quantum computing reports never ceases. This week – IonQ and Goldman Sachs tackle Monte Carlo on quantum hardware, Cambridge Quantum pu Read more…

Asetek Announces It Is Exiting HPC to Protect Future Profitability

September 22, 2021

Liquid cooling specialist Asetek, well-known in HPC circles for its direct-to-chip cooling technology that is inside some of the fastest supercomputers in the world, announced today that it is exiting the HPC space amid multiple supply chain issues related to the pandemic. Although pandemic supply chain... Read more…

TACC Supercomputer Delves Into Protein Interactions

September 22, 2021

Adenosine triphosphate (ATP) is a compound used to funnel energy from mitochondria to other parts of the cell, enabling energy-driven functions like muscle contractions. For ATP to flow, though, the interaction between the hexokinase-II (HKII) enzyme and the proteins found in a specific channel on the mitochondria’s outer membrane. Now, simulations conducted on supercomputers at the Texas Advanced Computing Center (TACC) have simulated... Read more…

The Latest MLPerf Inference Results: Nvidia GPUs Hold Sway but Here Come CPUs and Intel

September 22, 2021

The latest round of MLPerf inference benchmark (v 1.1) results was released today and Nvidia again dominated, sweeping the top spots in the closed (apples-to-ap Read more…

Why HPC Storage Matters More Now Than Ever: Analyst Q&A

September 17, 2021

With soaring data volumes and insatiable computing driving nearly every facet of economic, social and scientific progress, data storage is seizing the spotlight. Hyperion Research analyst and noted storage expert Mark Nossokoff looks at key storage trends in the context of the evolving HPC (and AI) landscape... Read more…

Ahead of ‘Dojo,’ Tesla Reveals Its Massive Precursor Supercomputer

June 22, 2021

In spring 2019, Tesla made cryptic reference to a project called Dojo, a “super-powerful training computer” for video data processing. Then, in summer 2020, Tesla CEO Elon Musk tweeted: “Tesla is developing a [neural network] training computer called Dojo to process truly vast amounts of video data. It’s a beast! … A truly useful exaflop at de facto FP32.” Read more…

Enter Dojo: Tesla Reveals Design for Modular Supercomputer & D1 Chip

August 20, 2021

Two months ago, Tesla revealed a massive GPU cluster that it said was “roughly the number five supercomputer in the world,” and which was just a precursor to Tesla’s real supercomputing moonshot: the long-rumored, little-detailed Dojo system. “We’ve been scaling our neural network training compute dramatically over the last few years,” said Milan Kovac, Tesla’s director of autopilot engineering. Read more…

Esperanto, Silicon in Hand, Champions the Efficiency of Its 1,092-Core RISC-V Chip

August 27, 2021

Esperanto Technologies made waves last December when it announced ET-SoC-1, a new RISC-V-based chip aimed at machine learning that packed nearly 1,100 cores onto a package small enough to fit six times over on a single PCIe card. Now, Esperanto is back, silicon in-hand and taking aim... Read more…

CentOS Replacement Rocky Linux Is Now in GA and Under Independent Control

June 21, 2021

The Rocky Enterprise Software Foundation (RESF) is announcing the general availability of Rocky Linux, release 8.4, designed as a drop-in replacement for the soon-to-be discontinued CentOS. The GA release is launching six-and-a-half months after Red Hat deprecated its support for the widely popular, free CentOS server operating system. The Rocky Linux development effort... Read more…

Intel Completes LLVM Adoption; Will End Updates to Classic C/C++ Compilers in Future

August 10, 2021

Intel reported in a blog this week that its adoption of the open source LLVM architecture for Intel’s C/C++ compiler is complete. The transition is part of In Read more…

Hot Chips: Here Come the DPUs and IPUs from Arm, Nvidia and Intel

August 25, 2021

The emergence of data processing units (DPU) and infrastructure processing units (IPU) as potentially important pieces in cloud and datacenter architectures was Read more…

AMD-Xilinx Deal Gains UK, EU Approvals — China’s Decision Still Pending

July 1, 2021

AMD’s planned acquisition of FPGA maker Xilinx is now in the hands of Chinese regulators after needed antitrust approvals for the $35 billion deal were receiv Read more…

Google Launches TPU v4 AI Chips

May 20, 2021

Google CEO Sundar Pichai spoke for only one minute and 42 seconds about the company’s latest TPU v4 Tensor Processing Units during his keynote at the Google I Read more…

Leading Solution Providers

Contributors

HPE Wins $2B GreenLake HPC-as-a-Service Deal with NSA

September 1, 2021

In the heated, oft-contentious, government IT space, HPE has won a massive $2 billion contract to provide HPC and AI services to the United States’ National Security Agency (NSA). Following on the heels of the now-canceled $10 billion JEDI contract (reissued as JWCC) and a $10 billion... Read more…

10nm, 7nm, 5nm…. Should the Chip Nanometer Metric Be Replaced?

June 1, 2020

The biggest cool factor in server chips is the nanometer. AMD beating Intel to a CPU built on a 7nm process node* – with 5nm and 3nm on the way – has been i Read more…

Julia Update: Adoption Keeps Climbing; Is It a Python Challenger?

January 13, 2021

The rapid adoption of Julia, the open source, high level programing language with roots at MIT, shows no sign of slowing according to data from Julialang.org. I Read more…

Quantum Roundup: IBM, Rigetti, Phasecraft, Oxford QC, China, and More

July 13, 2021

IBM yesterday announced a proof for a quantum ML algorithm. A week ago, it unveiled a new topology for its quantum processors. Last Friday, the Technical Univer Read more…

Intel Launches 10nm ‘Ice Lake’ Datacenter CPU with Up to 40 Cores

April 6, 2021

The wait is over. Today Intel officially launched its 10nm datacenter CPU, the third-generation Intel Xeon Scalable processor, codenamed Ice Lake. With up to 40 Read more…

Frontier to Meet 20MW Exascale Power Target Set by DARPA in 2008

July 14, 2021

After more than a decade of planning, the United States’ first exascale computer, Frontier, is set to arrive at Oak Ridge National Laboratory (ORNL) later this year. Crossing this “1,000x” horizon required overcoming four major challenges: power demand, reliability, extreme parallelism and data movement. Read more…

Intel Unveils New Node Names; Sapphire Rapids Is Now an ‘Intel 7’ CPU

July 27, 2021

What's a preeminent chip company to do when its process node technology lags the competition by (roughly) one generation, but outmoded naming conventions make it seem like it's two nodes behind? For Intel, the response was to change how it refers to its nodes with the aim of better reflecting its positioning within the leadership semiconductor manufacturing space. Intel revealed its new node nomenclature, and... Read more…

The Latest MLPerf Inference Results: Nvidia GPUs Hold Sway but Here Come CPUs and Intel

September 22, 2021

The latest round of MLPerf inference benchmark (v 1.1) results was released today and Nvidia again dominated, sweeping the top spots in the closed (apples-to-ap Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire