,

Dando un salto cuántico, la inteligencia artificial (IA) es una tecnología clave para la industria automotriz

Dando un salto cuántico, la inteligencia artificial (IA) es una tecnología clave para la industria automotriz

Cada vez más funciones de los vehículos se basan en la inteligencia artificial. Sin embargo, los procesadores convencionales e incluso los chips gráficos están llegando cada vez más a sus límites en lo que respecta a los cálculos necesarios para las redes neuronales. Porsche Engineering informa sobre nuevas tecnologías que acelerarán los cálculos de IA en el futuro.

La inteligencia artificial (IA) es una tecnología clave para la industria automotriz, y el hardware rápido es igualmente importante para los complejos cálculos de back-end involucrados. Después de todo, en el futuro solo será posible llevar nuevas funciones a la producción en serie con computadoras de alto rendimiento. “La conducción autónoma es una de las aplicaciones de IA más exigentes de todas”, explica el Dr. Joachim Schaper, Gerente Senior de IA y Big Data en Porsche Engineering. “Los algoritmos aprenden de una multitud de ejemplos recopilados por vehículos de prueba que utilizan cámaras, radares u otros sensores en el tráfico real”.

Dr. Joachim Schaper, gerente sénior de IA y Big Data en Porsche Engineering

dr. Joachim Schaper, Gerente Senior de IA y Big Data en Porsche Engineer

Los centros de datos convencionales son cada vez más incapaces de hacer frente a las crecientes demandas. “Ahora lleva días entrenar una sola variante de una red neuronal”, explica Schaper. Entonces, en su opinión, una cosa está clara: los fabricantes de automóviles necesitan nuevas tecnologías para los cálculos de IA que puedan ayudar a que los algoritmos aprendan mucho más rápido. Para lograr esto, se deben ejecutar en paralelo tantas multiplicaciones de matriz vectorial como sea posible en las complejas redes neuronales profundas (DNN), una tarea en la que se especializan las unidades de procesamiento de gráficos (GPU). Sin ellos, los increíbles avances en IA de los últimos años no habrían sido posibles.

50 veces el tamaño de una GPU

Sin embargo, las tarjetas gráficas no se diseñaron originalmente para el uso de IA, sino para procesar datos de imagen de la manera más eficiente posible. Están cada vez más al límite cuando se trata de algoritmos de entrenamiento para la conducción autónoma. Por lo tanto, se requiere hardware especializado en IA para cálculos aún más rápidos. La empresa californiana Cerebras ha presentado una posible solución. Su Wafer Scale Engine (WSE) se adapta de manera óptima a los requisitos de las redes neuronales al combinar la mayor potencia informática posible en un chip de computadora gigante. Es más de 50 veces el tamaño de un procesador de gráficos normal y ofrece espacio para 850 000 núcleos informáticos, más de 100 veces más que en una GPU superior actual.

Además, los ingenieros de Cerebras han conectado en red los núcleos computacionales junto con líneas de datos de gran ancho de banda. Según el fabricante, la red del Wafer Scale Engine transporta 220 petabits por segundo. Cerebras también ha ampliado el cuello de botella dentro de las GPU: los datos viajan entre la memoria y la unidad de cómputo casi 10 000 veces más rápido que en las GPU de alto rendimiento, a 20 petabytes por segundo.

 

Chip gigante: el Wafer Scale Engine de Cerebras combina una enorme potencia informática en un solo circuito integrado con una longitud lateral de más de 20 centímetros.

Chip gigante: el Wafer Scale Engine de Cerebras combina una enorme potencia informática en un solo circuito integrado con una longitud lateral de más de 20 centímetros.

Para ahorrar aún más tiempo, Cerebras imita un truco del cerebro. Allí, las neuronas funcionan solo cuando reciben señales de otras neuronas. Las muchas conexiones que están actualmente inactivas no necesitan ningún recurso. En las DNN, por otro lado, la multiplicación de matriz vectorial a menudo implica multiplicar por el número cero. Esto cuesta tiempo innecesariamente. Por lo tanto, Wafer Scale Engine se abstiene de hacerlo. “Todos los ceros se filtran”, escribe Cerebras en su libro blanco sobre el WSE. Entonces, el chip solo realiza operaciones que producen un resultado distinto de cero.

Un inconveniente del chip es su alto requerimiento de energía eléctrica de 23 kW y requiere refrigeración por agua. Por lo tanto, Cerebras ha desarrollado su propia carcasa de servidor para su uso en centros de datos. El Wafer Scale Engine ya se está probando en los centros de datos de algunos institutos de investigación. El experto en inteligencia artificial Joachim Schaper cree que el chip gigante de California también podría acelerar el desarrollo automotriz. “Al usar este chip, el entrenamiento de una semana podría reducirse teóricamente a unas pocas horas”, estima. “Sin embargo, la tecnología aún tiene que demostrarlo en pruebas prácticas”.

Luz en lugar de electrones

A pesar de lo inusual que es el nuevo chip, al igual que sus predecesores convencionales, también funciona con transistores convencionales. Empresas como Lightelligence y Lightmatter, con sede en Boston, quieren utilizar el medio de la luz mucho más rápido para los cálculos de IA en lugar de la electrónica comparativamente lenta, y están construyendo chips ópticos para hacerlo. Por lo tanto, los DNN podrían funcionar “al menos varios cientos de veces más rápido que los electrónicos”, escriben los desarrolladores de Lightelligence.

“Con Wafer Scale Engine, una semana de entrenamiento teóricamente podría reducirse a solo unas pocas horas”. Dr. Joachim Schaper, gerente sénior de IA y Big Data en Porsche Engineering

Para ello, Lightelligence y Lightmatter utilizan el fenómeno de la interferencia. Cuando las ondas de luz se amplifican o anulan entre sí, forman un patrón claro-oscuro. Si dirige la interferencia de cierta manera, el nuevo patrón corresponde a la multiplicación vector-matriz del patrón anterior. Entonces, las ondas de luz pueden “hacer matemáticas”. Para que esto sea práctico, los desarrolladores de Boston grabaron diminutas guías de luz en un chip de silicio. Como en un tejido textil, se cruzan varias veces. La interferencia tiene lugar en los cruces. En el medio, diminutos elementos calefactores regulan el índice de refracción de la guía de luz, lo que permite que las ondas de luz se desplacen entre sí. Esto permite controlar su interferencia y realizar multiplicaciones vector-matriz.

Sin embargo, las empresas de Boston no prescinden por completo de la electrónica. Combinan sus computadoras livianas con componentes electrónicos convencionales que almacenan datos y realizan todos los cálculos, excepto las multiplicaciones de vectores y matrices. Estos incluyen, por ejemplo, las funciones de activación no lineal que modifican los valores de salida de cada neurona antes de pasar a la siguiente capa.

Computación con luz: el chip Envise de Lightmatter utiliza fotones en lugar de electrones para calcular redes neuronales.  Los datos de entrada y salida son suministrados y recibidos por electrónica convencional.

Computación con luz: el chip Envise de Lightmatter utiliza fotones en lugar de electrones para calcular redes neuronales. Los datos de entrada y salida son suministrados y recibidos por electrónica convencional.

Con la combinación de computación óptica y digital, los DNN se pueden calcular extremadamente rápido. “Su principal ventaja es la baja latencia”, explica Lindsey Hunt, portavoz de Lightelligence. Por ejemplo, esto permite que la DNN detecte objetos en imágenes más rápido, como peatones y usuarios de scooters eléctricos. En la conducción autónoma, esto podría dar lugar a reacciones más rápidas en situaciones críticas. “Además, el sistema óptico toma más decisiones por vatio de energía eléctrica”, dijo Hunt. Eso es especialmente importante ya que el aumento de la potencia informática en los vehículos se produce cada vez más a expensas de la economía de combustible y la autonomía.

Las soluciones de Lightmatter y Lightelligence se pueden insertar como módulos en computadoras convencionales para acelerar los cálculos de IA, al igual que las tarjetas gráficas. En principio, también podrían integrarse en vehículos, por ejemplo, para implementar funciones de conducción autónoma. “Nuestra tecnología es muy adecuada para servir como motor de inferencia para un automóvil autónomo”, explica Lindsey Hunt. El experto en inteligencia artificial Schaper tiene una opinión similar: “Si Lightelligence tiene éxito en la construcción de componentes adecuados para automóviles, esto podría acelerar en gran medida la introducción de funciones complejas de inteligencia artificial en los vehículos”. La tecnología ya está lista para el mercado: la compañía está planeando sus primeras pruebas piloto con clientes en el año 2022.

La computadora cuántica como un turbo AI

Las computadoras cuánticas están algo más alejadas de la aplicación práctica. Ellos también acelerarán los cálculos de IA porque pueden procesar grandes cantidades de datos en paralelo. Para ello, trabajan con los llamados “qubits”. A diferencia de la unidad de información clásica, el bit, un qubit puede representar los dos valores binarios 0 y 1 simultáneamente. Los dos números coexisten en un estado de superposición que solo es posible en la mecánica cuántica.

“Cuanto más complicados son los patrones, más dificultad tienen las computadoras convencionales para distinguir clases”. Heike Riel, directora de IBM Research Quantum Europa/África

Las computadoras cuánticas podrían impulsar la inteligencia artificial cuando se trata de clasificar cosas, por ejemplo, en el tráfico. Hay muchas categorías diferentes de objetos allí, incluidas bicicletas, automóviles, peatones, señales, carreteras secas y mojadas. Difieren en términos de muchas propiedades, razón por la cual los expertos hablan de “reconocimiento de patrones en espacios de dimensiones superiores”.

“Cuanto más complicados son los patrones, más difícil es para las computadoras convencionales distinguir las clases”, explica Heike Riel, quien dirige la investigación cuántica de IBM en Europa y África. Eso se debe a que con cada dimensión, se vuelve más costoso calcular la similitud de dos objetos: ¿Qué tan similares son un conductor de e-scooter y un usuario de andador tratando de cruzar la calle? Las computadoras cuánticas pueden funcionar de manera eficiente en espacios de alta dimensión en comparación con las computadoras convencionales. Para ciertos problemas, esta propiedad podría ser útil y dar como resultado que algunos problemas se resuelvan más rápido con la ayuda de las computadoras cuánticas que con las computadoras convencionales de alto rendimiento.

Heike Riel, directora de IBM Research Quantum Europa/África

Heike Riel, directora de IBM Research Quantum Europa/África

Los investigadores de IBM han analizado modelos estadísticos que se pueden entrenar para la clasificación de datos. Los resultados iniciales sugieren que los modelos cuánticos inteligentemente elegidos funcionan mejor que los métodos convencionales para ciertos conjuntos de datos. Los modelos cuánticos son más fáciles de entrenar y parecen tener una mayor capacidad, lo que les permite aprender relaciones más complicadas.

Riel admite que, si bien las computadoras cuánticas actuales se pueden usar para probar estos algoritmos, aún no tienen una ventaja sobre las computadoras convencionales. Sin embargo, el desarrollo de las computadoras cuánticas avanza rápidamente. Tanto el número de qubits como su calidad aumentan constantemente. Otro factor importante es la velocidad, medida en operaciones de capa de circuito por segundo (CLOPS). Este número indica cuántos circuitos cuánticos pueden ejecutarse en la computadora cuántica por vez. Es uno de los tres criterios de rendimiento importantes de una computadora cuántica: escalabilidad, calidad y velocidad.

En un futuro previsible, debería ser posible demostrar la superioridad de las computadoras cuánticas para ciertas aplicaciones, es decir, que resuelven problemas de manera más rápida, eficiente y precisa que una computadora convencional. Pero la construcción de una computadora cuántica potente, con errores corregidos y de propósito general aún llevará algún tiempo. Los expertos estiman que llevará al menos otros diez años. Pero la espera podría valer la pena. Al igual que los chips ópticos o las nuevas arquitecturas para computadoras electrónicas, las computadoras cuánticas podrían ser la clave de la movilidad del futuro.

En breve

Cuando se trata de cálculos de IA, no solo los microprocesadores convencionales, sino también los chips gráficos, ahora están llegando a sus límites. Por lo tanto, empresas e investigadores de todo el mundo están trabajando en nuevas soluciones. Los chips en formato oblea y los ordenadores ligeros están cerca de hacerse realidad. En unos años, estos podrían complementarse con computadoras cuánticas para cálculos particularmente exigentes.

1
COMPARTE
,

Taking a quantum leap

Taking a quantum leap

More and more vehicle functions are based on artificial intelligence. However, conventional processors and even graphics chips are increasingly reaching their limits when it comes to calculations required for neural networks.  Porsche Engineering reports on new technologies that will speed up AI calculations in the future.

 

Artificial intelligence (AI) is a key technology for the automotive industry—and fast hardware is correspondingly important for the complex back-end calculations involved. After all, it will only be possible to bring new functions into series production in the future with high-performance computers. “Autonomous driving is one of the most demanding AI applications of all,” explains Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering. “The algorithms learn from a multitude of examples collected by test vehicles using cameras, radar, or other sensors in real traffic.”

Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering

Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineerin

Conventional data centers are increasingly unable to cope with the growing demands. “It now takes days to train a single variant of a neural network,” explains Schaper. So in his view, one thing is clear: Car manufacturers need new technologies for AI calculations that can help the algorithms learn much faster. To achieve this, as many vector-matrix multiplications as possible must be executed in parallel in the complex deep neural networks (DNNs)—a task in which graphics processing units (GPUs) specialize. Without them, the amazing advances in AI in recent years would not have been possible.

50 times the size of a GPU

Graphics cards were not originally designed for AI use, however, but to process image data as efficiently as possible. They are increasingly stretched to the limit when it comes to training algorithms for autonomous driving. Hardware specialized in AI is therefore required for even faster calculations. The Californian company Cerebras has presented a possible solution. Their Wafer Scale Engine (WSE) is optimally tailored to the requirements of neural networks by combining as much computing power as possible on one giant computer chip. It is more than 50 times the size of a normal graphics processor and offers space for 850,000 computing cores—over 100 times as many as on a current top GPU.

In addition, Cerebras engineers have networked the computational cores together with high-bandwidth data lines. According to the manufacturer, the network on the Wafer Scale Engine transports 220 petabits per second. Cerebras has also widened the bottleneck within the GPUs: Data travels between memory and computing unit nearly 10,000 times faster than in high-performance GPUs—at 20 petabytes per second.

 

Giant chip: Cerebras’ Wafer Scale Engine combines enormous computing power on a single integrated circuit with a side length of more than 20 centimeters.

Giant chip: Cerebras’ Wafer Scale Engine combines enormous computing power on a single integrated circuit with a side length of more than 20 centimeters.

To save even more time, Cerebras mimics a trick of the brain. There, neurons work only when they get signals from other neurons. The many connections that are currently inactive do not need any resources. In DNNs, on the other hand, vector-matrix multiplication often involves multiplying by the number zero. This costs time unnecessarily. The Wafer Scale Engine therefore refrains from doing so. “All zeros are filtered out,” Cerebras writes in its white paper on the WSE. So the chip only performs operations that produce a non-zero result.

One drawback of the chip is its high electrical power requirement of 23 kW and requires water cooling. Cerebras has therefore developed its own server housing for use in data centers. The Wafer Scale Engine is already being tested in the data centers of some research institutes. AI expert Joachim Schaper believes the giant chip from California could also accelerate automotive development. “By using this chip, a week’s training could theoretically be reduced to just a few hours,” he estimates. “However, the technology has yet to prove that in practical tests.”

Light instead of electrons

As unusual as the new chip is, like its conventional predecessors it also works with conventional transistors. Companies like Boston-based Lightelligence and Lightmatter want to use the much faster medium of light for AI calculations instead of comparatively slow electronics, and are building optical chips to do so. DNNs could thus work “at least several hundred times faster than electronic ones,” write developers at Lightelligence.

“With the Wafer Scale Engine, a week of training could theoretically be reduced to just a few hours.”Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering

To do this, Lightelligence and Lightmatter use the phenomenon of interference. When light waves amplify or cancel each other, they form a light-dark pattern. If you direct the interference in a certain way, the new pattern corresponds to the vector-matrix multiplication of the old pattern. So the light waves can “do math.” To make this practical, the Boston developers etched tiny light guides into a silicon chip. Like in a textile fabric, they cross each other several times. Interference takes place at the crossings. In between, tiny heating elements regulate the refractive index of the light guide, allowing the light waves to be shifted against each other. This makes it possible to control their interference and perform vector-matrix multiplications.

However, the Boston companies do not dispense with electronics altogether. They combine their light computers with conventional electronic components that store data and perform all calculations except vector-matrix multiplications. These include, for example, the nonlinear activation functions that modify the output values of each neuron before they move on to the next layer.

Computing with light: Lightmatter’s Envise chip uses photons instead of electrons to calculate neural networks. The input and output data is supplied and received by conventional electronics.

Computing with light: Lightmatter’s Envise chip uses photons instead of electrons to calculate neural networks. The input and output data is supplied and received by conventional electronics.

With the combination of optical and digital computing, DNNs can be computed extremely quickly. “Their main advantage is low latency,” explains Lindsey Hunt, a spokesperson for Lightelligence. For example, this allows the DNN to detect objects in images faster, such as pedestrians and e-scooter riders. In autonomous driving, this could lead to faster reactions in critical situations. “In addition, the optical system makes more decisions per watt of electrical energy,” Hunt said. That’s especially important as increasing computing power in vehicles increasingly comes at the expense of fuel economy and range.

The solutions from Lightmatter and Lightelligence can be inserted as modules into conventional computers to speed up AI computations—much like graphics cards. In principle, they could also be integrated into  vehicles, for example to implement autonomous driving functions. “Our technology is very well suited to serve as an inference engine for an autonomous car,” explains Lindsey Hunt. AI expert Schaper has a similar view: “If Lightelligence succeeds in building components suitable for automobiles, this could greatly accelerate the introduction of complex AI functions in vehicles.” The technology is now ready for the market: The company is planning its first pilot tests with customers in the year 2022.

The quantum computer as an AI turbo

Quantum computers are somewhat further away from practical application. They, too, will accelerate AI calculations because they can process vast amounts of data in parallel. To do this, they work with so-called “qubits.” Unlike the classical unit of information, the bit, a qubit can represent the two binary values 0 and 1 simultaneously. The two numbers coexist in a superposition state that is only possible in quantum mechanics.

“The more complicated the patterns, the more difficulty conventional computers have distinguishing classes.”Heike Riel, Lead IBM Research Quantum Europe/Africa

Quantum computers could turbocharge artificial intelligence when if comes to classifying things, for example in traffic. There are many different categories of objects there, including bicycles, cars, pedestrians, signs, wet and dry roads. They differ in terms of many properties, which is why experts talk about “pattern recognition in higher-dimensional spaces.”

“The more complicated the patterns, the harder it is for conventional computers to distinguish classes,” explains Heike Riel, who heads IBM’s quantum research in Europe and Africa. That’s because with each dimension, it becomes more costly to calculate the similarity of two objects: How similar are an e-scooter rider and a rollator user trying to cross the street? Quantum computers can work efficiently in high-dimensional spaces compared to conventional computers. For certain problems, this property could be useful and result in some problems being solved faster with the help of quantum computers than with conventional high-performance computers.

Heike Riel, Lead IBM Research Quantum Europe/Africa

Heike Riel, Lead IBM Research Quantum Europe/Africa

IBM researchers have analyzed statistical models that can be trained for data classification. Initial results suggest that cleverly chosen quantum models work better than conventional methods for certain datasets. The quantum models are easier to train and appear to have greater capacity—allowing them to learn more complicated relationships.

Riel admits that while today’s quantum computers can be used to test these algorithms, they do not yet have an advantage over conventional computers. However, the development of quantum computers is progressing rapidly. Both the number of qubits and their quality are steadily increasing. Another important factor is speed, measured in Circuit Layer Operations per Second (CLOPS). This number denotes how many quantum circuits can run on the quantum computer per time. It is one of the three important performance criteria of a quantum computer: scalability, quality, and speed.

In the foreseeable future, it should be possible to demonstrate the superiority of quantum computers for certain applications—that is, that they solve problems faster, more efficiently, and more precisely than a conventional computer. But building a powerful, error-corrected, general-purpose quantum computer will still take some time. Experts estimate that it will take at least another ten years. But the wait could be worth it. Like optical chips or new architectures for electronic computers, quantum computers could be the key to the mobility of the future.

In brief

When it comes to AI calculations, not only conventional microprocessors, but also graphics chips, are now reaching their limits. Companies and researchers worldwide are therefore working on new solutions. Chips in wafer format and light computers are close to becoming reality. In a few years, these could be supplemented by quantum computers for particularly demanding calculations.

,

Taking a quantum leap

Taking a quantum leap

More and more vehicle functions are based on artificial intelligence. However, conventional processors and even graphics chips are increasingly reaching their limits when it comes to calculations required for neural networks.  Porsche Engineering reports on new technologies that will speed up AI calculations in the future.

 

Artificial intelligence (AI) is a key technology for the automotive industry—and fast hardware is correspondingly important for the complex back-end calculations involved. After all, it will only be possible to bring new functions into series production in the future with high-performance computers. “Autonomous driving is one of the most demanding AI applications of all,” explains Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering. “The algorithms learn from a multitude of examples collected by test vehicles using cameras, radar, or other sensors in real traffic.”

Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering

Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineerin

Conventional data centers are increasingly unable to cope with the growing demands. “It now takes days to train a single variant of a neural network,” explains Schaper. So in his view, one thing is clear: Car manufacturers need new technologies for AI calculations that can help the algorithms learn much faster. To achieve this, as many vector-matrix multiplications as possible must be executed in parallel in the complex deep neural networks (DNNs)—a task in which graphics processing units (GPUs) specialize. Without them, the amazing advances in AI in recent years would not have been possible.

50 times the size of a GPU

Graphics cards were not originally designed for AI use, however, but to process image data as efficiently as possible. They are increasingly stretched to the limit when it comes to training algorithms for autonomous driving. Hardware specialized in AI is therefore required for even faster calculations. The Californian company Cerebras has presented a possible solution. Their Wafer Scale Engine (WSE) is optimally tailored to the requirements of neural networks by combining as much computing power as possible on one giant computer chip. It is more than 50 times the size of a normal graphics processor and offers space for 850,000 computing cores—over 100 times as many as on a current top GPU.

In addition, Cerebras engineers have networked the computational cores together with high-bandwidth data lines. According to the manufacturer, the network on the Wafer Scale Engine transports 220 petabits per second. Cerebras has also widened the bottleneck within the GPUs: Data travels between memory and computing unit nearly 10,000 times faster than in high-performance GPUs—at 20 petabytes per second.

 

Giant chip: Cerebras’ Wafer Scale Engine combines enormous computing power on a single integrated circuit with a side length of more than 20 centimeters.

Giant chip: Cerebras’ Wafer Scale Engine combines enormous computing power on a single integrated circuit with a side length of more than 20 centimeters.

To save even more time, Cerebras mimics a trick of the brain. There, neurons work only when they get signals from other neurons. The many connections that are currently inactive do not need any resources. In DNNs, on the other hand, vector-matrix multiplication often involves multiplying by the number zero. This costs time unnecessarily. The Wafer Scale Engine therefore refrains from doing so. “All zeros are filtered out,” Cerebras writes in its white paper on the WSE. So the chip only performs operations that produce a non-zero result.

One drawback of the chip is its high electrical power requirement of 23 kW and requires water cooling. Cerebras has therefore developed its own server housing for use in data centers. The Wafer Scale Engine is already being tested in the data centers of some research institutes. AI expert Joachim Schaper believes the giant chip from California could also accelerate automotive development. “By using this chip, a week’s training could theoretically be reduced to just a few hours,” he estimates. “However, the technology has yet to prove that in practical tests.”

Light instead of electrons

As unusual as the new chip is, like its conventional predecessors it also works with conventional transistors. Companies like Boston-based Lightelligence and Lightmatter want to use the much faster medium of light for AI calculations instead of comparatively slow electronics, and are building optical chips to do so. DNNs could thus work “at least several hundred times faster than electronic ones,” write developers at Lightelligence.

“With the Wafer Scale Engine, a week of training could theoretically be reduced to just a few hours.”Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering

To do this, Lightelligence and Lightmatter use the phenomenon of interference. When light waves amplify or cancel each other, they form a light-dark pattern. If you direct the interference in a certain way, the new pattern corresponds to the vector-matrix multiplication of the old pattern. So the light waves can “do math.” To make this practical, the Boston developers etched tiny light guides into a silicon chip. Like in a textile fabric, they cross each other several times. Interference takes place at the crossings. In between, tiny heating elements regulate the refractive index of the light guide, allowing the light waves to be shifted against each other. This makes it possible to control their interference and perform vector-matrix multiplications.

However, the Boston companies do not dispense with electronics altogether. They combine their light computers with conventional electronic components that store data and perform all calculations except vector-matrix multiplications. These include, for example, the nonlinear activation functions that modify the output values of each neuron before they move on to the next layer.

Computing with light: Lightmatter’s Envise chip uses photons instead of electrons to calculate neural networks. The input and output data is supplied and received by conventional electronics.

Computing with light: Lightmatter’s Envise chip uses photons instead of electrons to calculate neural networks. The input and output data is supplied and received by conventional electronics.

With the combination of optical and digital computing, DNNs can be computed extremely quickly. “Their main advantage is low latency,” explains Lindsey Hunt, a spokesperson for Lightelligence. For example, this allows the DNN to detect objects in images faster, such as pedestrians and e-scooter riders. In autonomous driving, this could lead to faster reactions in critical situations. “In addition, the optical system makes more decisions per watt of electrical energy,” Hunt said. That’s especially important as increasing computing power in vehicles increasingly comes at the expense of fuel economy and range.

The solutions from Lightmatter and Lightelligence can be inserted as modules into conventional computers to speed up AI computations—much like graphics cards. In principle, they could also be integrated into  vehicles, for example to implement autonomous driving functions. “Our technology is very well suited to serve as an inference engine for an autonomous car,” explains Lindsey Hunt. AI expert Schaper has a similar view: “If Lightelligence succeeds in building components suitable for automobiles, this could greatly accelerate the introduction of complex AI functions in vehicles.” The technology is now ready for the market: The company is planning its first pilot tests with customers in the year 2022.

The quantum computer as an AI turbo

Quantum computers are somewhat further away from practical application. They, too, will accelerate AI calculations because they can process vast amounts of data in parallel. To do this, they work with so-called “qubits.” Unlike the classical unit of information, the bit, a qubit can represent the two binary values 0 and 1 simultaneously. The two numbers coexist in a superposition state that is only possible in quantum mechanics.

“The more complicated the patterns, the more difficulty conventional computers have distinguishing classes.”Heike Riel, Lead IBM Research Quantum Europe/Africa

Quantum computers could turbocharge artificial intelligence when if comes to classifying things, for example in traffic. There are many different categories of objects there, including bicycles, cars, pedestrians, signs, wet and dry roads. They differ in terms of many properties, which is why experts talk about “pattern recognition in higher-dimensional spaces.”

“The more complicated the patterns, the harder it is for conventional computers to distinguish classes,” explains Heike Riel, who heads IBM’s quantum research in Europe and Africa. That’s because with each dimension, it becomes more costly to calculate the similarity of two objects: How similar are an e-scooter rider and a rollator user trying to cross the street? Quantum computers can work efficiently in high-dimensional spaces compared to conventional computers. For certain problems, this property could be useful and result in some problems being solved faster with the help of quantum computers than with conventional high-performance computers.

Heike Riel, Lead IBM Research Quantum Europe/Africa

Heike Riel, Lead IBM Research Quantum Europe/Africa

IBM researchers have analyzed statistical models that can be trained for data classification. Initial results suggest that cleverly chosen quantum models work better than conventional methods for certain datasets. The quantum models are easier to train and appear to have greater capacity—allowing them to learn more complicated relationships.

Riel admits that while today’s quantum computers can be used to test these algorithms, they do not yet have an advantage over conventional computers. However, the development of quantum computers is progressing rapidly. Both the number of qubits and their quality are steadily increasing. Another important factor is speed, measured in Circuit Layer Operations per Second (CLOPS). This number denotes how many quantum circuits can run on the quantum computer per time. It is one of the three important performance criteria of a quantum computer: scalability, quality, and speed.

In the foreseeable future, it should be possible to demonstrate the superiority of quantum computers for certain applications—that is, that they solve problems faster, more efficiently, and more precisely than a conventional computer. But building a powerful, error-corrected, general-purpose quantum computer will still take some time. Experts estimate that it will take at least another ten years. But the wait could be worth it. Like optical chips or new architectures for electronic computers, quantum computers could be the key to the mobility of the future.

In brief

When it comes to AI calculations, not only conventional microprocessors, but also graphics chips, are now reaching their limits. Companies and researchers worldwide are therefore working on new solutions. Chips in wafer format and light computers are close to becoming reality. In a few years, these could be supplemented by quantum computers for particularly demanding calculations.

,

Taking a quantum leap

Taking a quantum leap

More and more vehicle functions are based on artificial intelligence. However, conventional processors and even graphics chips are increasingly reaching their limits when it comes to calculations required for neural networks.  Porsche Engineering reports on new technologies that will speed up AI calculations in the future.

 

Artificial intelligence (AI) is a key technology for the automotive industry—and fast hardware is correspondingly important for the complex back-end calculations involved. After all, it will only be possible to bring new functions into series production in the future with high-performance computers. “Autonomous driving is one of the most demanding AI applications of all,” explains Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering. “The algorithms learn from a multitude of examples collected by test vehicles using cameras, radar, or other sensors in real traffic.”

Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering

Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineerin

Conventional data centers are increasingly unable to cope with the growing demands. “It now takes days to train a single variant of a neural network,” explains Schaper. So in his view, one thing is clear: Car manufacturers need new technologies for AI calculations that can help the algorithms learn much faster. To achieve this, as many vector-matrix multiplications as possible must be executed in parallel in the complex deep neural networks (DNNs)—a task in which graphics processing units (GPUs) specialize. Without them, the amazing advances in AI in recent years would not have been possible.

50 times the size of a GPU

Graphics cards were not originally designed for AI use, however, but to process image data as efficiently as possible. They are increasingly stretched to the limit when it comes to training algorithms for autonomous driving. Hardware specialized in AI is therefore required for even faster calculations. The Californian company Cerebras has presented a possible solution. Their Wafer Scale Engine (WSE) is optimally tailored to the requirements of neural networks by combining as much computing power as possible on one giant computer chip. It is more than 50 times the size of a normal graphics processor and offers space for 850,000 computing cores—over 100 times as many as on a current top GPU.

In addition, Cerebras engineers have networked the computational cores together with high-bandwidth data lines. According to the manufacturer, the network on the Wafer Scale Engine transports 220 petabits per second. Cerebras has also widened the bottleneck within the GPUs: Data travels between memory and computing unit nearly 10,000 times faster than in high-performance GPUs—at 20 petabytes per second.

 

Giant chip: Cerebras’ Wafer Scale Engine combines enormous computing power on a single integrated circuit with a side length of more than 20 centimeters.

Giant chip: Cerebras’ Wafer Scale Engine combines enormous computing power on a single integrated circuit with a side length of more than 20 centimeters.

To save even more time, Cerebras mimics a trick of the brain. There, neurons work only when they get signals from other neurons. The many connections that are currently inactive do not need any resources. In DNNs, on the other hand, vector-matrix multiplication often involves multiplying by the number zero. This costs time unnecessarily. The Wafer Scale Engine therefore refrains from doing so. “All zeros are filtered out,” Cerebras writes in its white paper on the WSE. So the chip only performs operations that produce a non-zero result.

One drawback of the chip is its high electrical power requirement of 23 kW and requires water cooling. Cerebras has therefore developed its own server housing for use in data centers. The Wafer Scale Engine is already being tested in the data centers of some research institutes. AI expert Joachim Schaper believes the giant chip from California could also accelerate automotive development. “By using this chip, a week’s training could theoretically be reduced to just a few hours,” he estimates. “However, the technology has yet to prove that in practical tests.”

Light instead of electrons

As unusual as the new chip is, like its conventional predecessors it also works with conventional transistors. Companies like Boston-based Lightelligence and Lightmatter want to use the much faster medium of light for AI calculations instead of comparatively slow electronics, and are building optical chips to do so. DNNs could thus work “at least several hundred times faster than electronic ones,” write developers at Lightelligence.

“With the Wafer Scale Engine, a week of training could theoretically be reduced to just a few hours.”Dr. Joachim Schaper, Senior Manager AI and Big Data at Porsche Engineering

To do this, Lightelligence and Lightmatter use the phenomenon of interference. When light waves amplify or cancel each other, they form a light-dark pattern. If you direct the interference in a certain way, the new pattern corresponds to the vector-matrix multiplication of the old pattern. So the light waves can “do math.” To make this practical, the Boston developers etched tiny light guides into a silicon chip. Like in a textile fabric, they cross each other several times. Interference takes place at the crossings. In between, tiny heating elements regulate the refractive index of the light guide, allowing the light waves to be shifted against each other. This makes it possible to control their interference and perform vector-matrix multiplications.

However, the Boston companies do not dispense with electronics altogether. They combine their light computers with conventional electronic components that store data and perform all calculations except vector-matrix multiplications. These include, for example, the nonlinear activation functions that modify the output values of each neuron before they move on to the next layer.

Computing with light: Lightmatter’s Envise chip uses photons instead of electrons to calculate neural networks. The input and output data is supplied and received by conventional electronics.

Computing with light: Lightmatter’s Envise chip uses photons instead of electrons to calculate neural networks. The input and output data is supplied and received by conventional electronics.

With the combination of optical and digital computing, DNNs can be computed extremely quickly. “Their main advantage is low latency,” explains Lindsey Hunt, a spokesperson for Lightelligence. For example, this allows the DNN to detect objects in images faster, such as pedestrians and e-scooter riders. In autonomous driving, this could lead to faster reactions in critical situations. “In addition, the optical system makes more decisions per watt of electrical energy,” Hunt said. That’s especially important as increasing computing power in vehicles increasingly comes at the expense of fuel economy and range.

The solutions from Lightmatter and Lightelligence can be inserted as modules into conventional computers to speed up AI computations—much like graphics cards. In principle, they could also be integrated into  vehicles, for example to implement autonomous driving functions. “Our technology is very well suited to serve as an inference engine for an autonomous car,” explains Lindsey Hunt. AI expert Schaper has a similar view: “If Lightelligence succeeds in building components suitable for automobiles, this could greatly accelerate the introduction of complex AI functions in vehicles.” The technology is now ready for the market: The company is planning its first pilot tests with customers in the year 2022.

The quantum computer as an AI turbo

Quantum computers are somewhat further away from practical application. They, too, will accelerate AI calculations because they can process vast amounts of data in parallel. To do this, they work with so-called “qubits.” Unlike the classical unit of information, the bit, a qubit can represent the two binary values 0 and 1 simultaneously. The two numbers coexist in a superposition state that is only possible in quantum mechanics.

“The more complicated the patterns, the more difficulty conventional computers have distinguishing classes.”Heike Riel, Lead IBM Research Quantum Europe/Africa

Quantum computers could turbocharge artificial intelligence when if comes to classifying things, for example in traffic. There are many different categories of objects there, including bicycles, cars, pedestrians, signs, wet and dry roads. They differ in terms of many properties, which is why experts talk about “pattern recognition in higher-dimensional spaces.”

“The more complicated the patterns, the harder it is for conventional computers to distinguish classes,” explains Heike Riel, who heads IBM’s quantum research in Europe and Africa. That’s because with each dimension, it becomes more costly to calculate the similarity of two objects: How similar are an e-scooter rider and a rollator user trying to cross the street? Quantum computers can work efficiently in high-dimensional spaces compared to conventional computers. For certain problems, this property could be useful and result in some problems being solved faster with the help of quantum computers than with conventional high-performance computers.

Heike Riel, Lead IBM Research Quantum Europe/Africa

Heike Riel, Lead IBM Research Quantum Europe/Africa

IBM researchers have analyzed statistical models that can be trained for data classification. Initial results suggest that cleverly chosen quantum models work better than conventional methods for certain datasets. The quantum models are easier to train and appear to have greater capacity—allowing them to learn more complicated relationships.

Riel admits that while today’s quantum computers can be used to test these algorithms, they do not yet have an advantage over conventional computers. However, the development of quantum computers is progressing rapidly. Both the number of qubits and their quality are steadily increasing. Another important factor is speed, measured in Circuit Layer Operations per Second (CLOPS). This number denotes how many quantum circuits can run on the quantum computer per time. It is one of the three important performance criteria of a quantum computer: scalability, quality, and speed.

In the foreseeable future, it should be possible to demonstrate the superiority of quantum computers for certain applications—that is, that they solve problems faster, more efficiently, and more precisely than a conventional computer. But building a powerful, error-corrected, general-purpose quantum computer will still take some time. Experts estimate that it will take at least another ten years. But the wait could be worth it. Like optical chips or new architectures for electronic computers, quantum computers could be the key to the mobility of the future.

In brief

When it comes to AI calculations, not only conventional microprocessors, but also graphics chips, are now reaching their limits. Companies and researchers worldwide are therefore working on new solutions. Chips in wafer format and light computers are close to becoming reality. In a few years, these could be supplemented by quantum computers for particularly demanding calculations.

0 replies

Leave a Reply

Want to join the discussion?
Feel free to contribute!

Leave a Reply