System Bits: Dec. 12

Increasing performance scaling with packageless processors
Demand for increasing performance is far outpacing the capability of traditional methods for performance scaling. Disruptive solutions are needed to advance beyond incremental improvements. Traditionally, processors reside inside packages to enable PCB-based integration. However, a team of researchers from the Department of Electrical and Computer Engineering from the University of California, Los Angeles along with colleagues from the Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign argue that packages reduce the potential memory bandwidth of a processor by at least one order of magnitude, allowable thermal design power (TDP) by up to 70%, and area efficiency by a factor of 5 to 18.

At the same time, silicon chips have scaled well while packages have not. The team is proposing packageless processors in which the packages have been removed and dies are directly mounted on a silicon board using a novel integration technology, Silicon Interconnection Fabric (Si-IF).

The researchers have shown that Si-IF-based packageless processors outperform their packaged counterparts by up to 58% (16% average), 136% (103% average), and 295% (80% average) due to increased memory bandwidth, increased allowable TDP, and reduced area respectively.

They have also extended the concept of packageless processing to the entire processor and memory system, where the area footprint reduction was up to 76%.

At first glance Si-IF technology seems similar to interposers, but the team asserted it is fundamentally different. Interposers use through-silicon-vias (TSV) and because of aspect ratio limitation of TSVs, the interposer needs to be thinned and thus it becomes fragile and size limited. In fact, interposers are typically limited to the maximum mask field size (e.g., ∼830 mm2 which is the same as maximum SoC size) to avoid stitching. Though larger interposers can be built using stitching, they are much costlier and have lower yield. Also, interposers need packages for mechanical support and for space transformation to accommodate larger I/O connections to the PCB. Therefore, connections with chips outside of the interposers continues to suffer from the issues of conventional packaging. On the other hand, Si-IF is a standalone rigid interconnect substrate capable of scaling up to a full size of a wafer and doesn’t require packages for mechanical support.

Illuminating the inner workings of language processing AI systems
Neural networks, which learn to perform computational tasks by analyzing huge sets of training data, have been responsible for the most impressive recent advances in artificial intelligence, including speech-recognition and automatic-translation systems. But during training, a neural net continually adjusts its internal settings in ways that even its creators can’t interpret. MIT researchers reminded that recent work in computer science has focused on clever techniques for determining just how neural nets do what they do. For instance, several recent papers by researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Qatar Computing Research Institute have used a recently developed interpretive technique, which had been applied in other areas, to analyze neural networks trained to do machine translation and speech recognition.

They found empirical support for some common intuitions about how the networks probably work. For example, they said systems seem to concentrate on lower-level tasks, such as sound recognition or part-of-speech recognition, before moving on to higher-level tasks, such as transcription or semantic interpretation.

Neural nets are so named because they roughly approximate the structure of the human brain. Typically, they’re arranged into layers, and each layer consists of many simple processing units — nodes — each of which is connected to several nodes in the layers above and below. Data is fed into the lowest layer, whose nodes process it and pass it to the next layer. The connections between layers have different “weights,” which determine how much the output of any one node figures into the calculation performed by the next.
Source: MIT

Interestingly, the researchers also found a surprising omission in the type of data the translation network considers, and they have shown that correcting that omission improves the network’s performance. The improvement is modest, but it points toward the possibility that analysis of neural networks could help improve the accuracy of artificial intelligence systems.

“In machine translation, historically, there was sort of a pyramid with different layers,” said Jim Glass, a CSAIL senior research scientist who worked on the project with Yonatan Belinkov, an MIT graduate student in electrical engineering and computer science. “At the lowest level there was the word, the surface forms, and the top of the pyramid was some kind of interlingual representation, and you’d have different layers where you were doing syntax, semantics. This was a very abstract notion, but the idea was the higher up you went in the pyramid, the easier it would be to translate to a new language, and then you’d go down again. So part of what Yonatan is doing is trying to figure out what aspects of this notion are being encoded in the network.”

The work on machine translation was presented recently in two papers at the International Joint Conference on Natural Language Processing. On one, Belinkov is first author, and Glass is senior author, and on the other, Belinkov is a co-author. On both, they’re joined by researchers from the Qatar Computing Research Institute (QCRI), including Lluís Màrquez, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and Stephan Vogel. Belinkov and Glass are sole authors on the paper analyzing speech recognition systems, which Belinkov presented at the Neural Information Processing Symposium recently.

Braided qubits could form component of topological quantum computer
Rice University physicists dedicated to creating the working components of a fault-tolerant quantum computer have succeeded in creating a previously unseen state of matter.

Rice’s “topological excitonic insulators” are made of sheets of semiconductors (top) that become insulators at a critical temperature around 10 kelvins. At the critical point, a superfluid quantum liquid of excitons — pairs of negatively charged electrons (blue dots) and positively charged electron holes (red dots) — forms inside the devices (bottom) and electricity ceases to pass through them.
Source: Rice University

The “topological excitonic insulator” was observed in tests at Rice by an international team from the United States and China. The researchers reported their findings recently in the journal Nature Communications.

The team said this device could potentially be used in a topological quantum computer, a type of quantum computer that stores information in quantum particles that are “braided” together like knots that are not easily broken. These stable, braided “topological” quantum bits, or topological qubits, could overcome one of the primary limitations of quantum computing today: Qubits that are nontopological easily “decohere” and lose the information they are storing.

Conventional computers use binary data, information that is stored as ones or zeros. Thanks to the quirks of quantum mechanics, qubits can represent both ones, zeros and a third state that’s both a one and a zero at the same time. This third state can be used to speed up computation, so much so that a quantum computer with just a few dozen qubits could finish some computations as quickly as a microchip with a billion binary transistors, the researchers reminded.

In the new study, Rice physicist Rui-Rui Du and former Rice graduate student Lingjie Du (no relation) collaborated with researchers from Rice, Peking University and the Chinese Academy of Sciences to create excitonic insulators made of tiny slivers of ultrapure, stacked semiconductors. The devices, which are no more than 100 microns wide, contain a sheet of indium arsenide atop a sheet of gallium antimony. When cooled in a bath of liquid helium to a critically low temperature around 10 kelvins, a superfluid quantum liquid forms inside the devices and electricity ceases to pass through them.

“This is very much like the process in a superconductor, where you have electrons that are attracted to one another to form pairs that flow without resistance,” said Rui-Rui Du, a professor of physics and astronomy at Rice and a researcher at the Rice Center for Quantum Materials (RCQM). “In our case, electrons pair with positively charged ‘electron holes’ to create a superfluid with a net charge of zero.”

Lingjie Du, now a postdoctoral researcher at Columbia University, said, “It’s a collective effect, so to an outside observer the system conducts electricity normally until it’s cooled to the critical temperature, where it suddenly changes phase to become a perfect insulator.”

To prove that the device was the long-sought excitonic insulator, the team first had to show the fluid was a quantum condensate. That task fell to Xinwei Li, a graduate student in the laboratory of RCQM researcher Junichiro Kono. Li and Kono, a professor of electrical and computer engineering at Rice, shined terahertz waves through the devices as they were cooled to the critical temperature and found that the samples absorbed terahertz energy in two distinct bands — a signature of quantum condensation.

Showing the device was topological involved testing for electrical conduction in a one-dimensional band around their perimeter.

“This novel property of the edge state is the thing that people are very interested in,” Rui-Rui Du said. “This edge state has no electrical resistance, and you get conduction in which electrons are tied to their spin moment. If they have one type of spin, they go clockwise and if they have the other they go counterclockwise.”

Braiding circuits built on these opposing electron streams would have inherent topological signatures that could be used to form fault-tolerant qubits.

“The other beauty of this is that the same principles still apply at room temperature,” Rui-Rui Du said. “There are atomically layered materials such as tungsten disulfide that could potentially be used to create this same effect at room temperature, provided they could be made in pure enough form.”