Figure 1: IPU accelerator card from the UK artificial intelligence chip hardware design startup GraphcoreHe uses black marker to drill and graph on the "nodes" of the human brain, which are usually responsible for "thinking or thinking" in the brain. His startups are trying to simulate these neurons and synapses in the next generation of computer processors, and the company bets that the next generation of computer processors can help "smart mechanization."
AI is often thought to be a complex software for mining a large number of data sets, but Noel and its co-founder, Nigel Toon, said the computer running the software still has a bigger obstacle. sitting in a well-ventilated office in the british port city of bristol, the problem is that the chip itself, based on their functionality, can be divided into a central processing unit cpu or a graphics processing unit gpu, They do not "think" in any identifiable manner of human-like.
The human brain uses intuition to simplify certain problems, such as identifying a friend who is approaching, and the computer may try to analyze each pixel of that person's face and compare it to a database containing billions of images, then Will try to say hello. This precision makes sense when the computer is primarily used as a calculator, but for AI it is very inefficient and consumes a lot of energy to process all relevant data.
In 2016, Knowles and the more business-savvy Tuen created Graphcore, which put "less accurate" computing at the heart of the chip, called the intelligent processing unit (IPU). "the concepts in your brain are quite vague," says Mr Knowles. It's actually a very similar collection of data points, allowing you to produce accurate ideas. " Knowles's English accent and frequent giggling led him to compare him to the dean of Hogwarts College in Harry Potter.
There are various theories about why human intelligence is formed in this way. But for machine learning systems, they need to deal with large and irregular unstructured information structures (ie graphics). To create a chip dedicated to connecting data points like brain nodes, it may be the key to the continued evolution of AI. Knowles said: "We want to build a high-performance computer that can handle numbers in a very inaccurate way."
In other words, Graphcore is developing a "brain" for computers, and if its co-founders' ideas are right, it will be able to process information more like humans, rather than falsifying information through large-scale numerical operations. Thun explained: "For decades, we have been telling the machine what to do for the camp, but now we are no longer doing this." He described how the Graphcore chip teaches machine learning: "It's like going back. In the 1970s, when the microprocessor was just released, we needed to completely renovateIntel. ”
Investor Herman Hauser (Hermann Hauser), co-founder of Arm Holdings Plc, controls the most widely used chip design. Hauser bet that Knowles and Tuen's IPU would set off the next wave of computing. "this has only happened three times in computer history," Hauser said. "it's the third IPU of GPU,Graphcore in the 1970s, CPU,20 in the 1990s."
Figure 2: IPU of the Graphcore officeserverframe
Graphcore originated from a series of seminars organized by Hauser in 2011 and 2012 at the Royal Society of Cambridge, a scientific group of alumni of Isaac Newton (Isaac Newton) and Charles Darwin (Charles Darwin). At King's College's luxury restaurant, AI experts, neuroscientists, statisticians and zoologists debate the impact of advanced computing technology on society.
Hauser believes that Knowles “has a earth-sized brain” and he feels uncomfortable in this “ivory tower” despite his career from Cambridge. After graduating in the 1980s, Knowles studied early neural networks in a research laboratory at the British government. Later, he co-founded the wireless processor startup Element 14, and sold it to Broadcom in 2000 for $640 million.
Soon after, Knowles and Thun, who had experience in semiconductor entrepreneurship, first cooperated. In 2002, they created mobile chip maker Icera and sold it to Nvidia for less than 10 years at $436 million. At the time, the two were not ready to retire, Thun said: "We are not good at playing golf." When Knowles went to the Cambridge University lecture series, they were discussing other ideas. Knowles recalled: "I am the sly guy in the room, wearing a chimney cap, just want to do something. You know: ‘Don’t worry about thermodynamics, I want to be a steam engine!’”
When Steve Young (Steve Young), a professor of information engineering at Cambridge University, gave a speech on calculating the limits of dialogue systems, Knowles kept asking him questions about energy efficiency. Steve Young later sold a voice processing service to Apple, which is now used for Siri. "I asked him about the accuracy of the numbers he used in the algorithm, which, in Steve's opinion, was a bit off the point," Knowles said. But he stressed that in silicon materials, "the accuracy of numbers as a determinant of energy is very important."
A few days later, Steve Young sent an email to Knowles saying his students had investigated the incident and found that they used 64-digit data for each calculation. They realized that they could perform the same function with 8-bit data, as Knowles suggested, but the operation was not so accurate. When a computer has fewer mathematical tasks to do, it can use the energy saved to process more numbers. It's a bit like the conversion of the human brain from computing the GPS coordinates of a restaurant to just remembering its name and neighbors.
"If we make a processor that is more suitable for this kind of work, we can increase the performance by a thousand times," said Steve Young and others, who decided that they had to create a Graphcore. As early as 2013, they started raising funds to develop the idea and presented the company to the world in 2016.
The semiconductor industry is currently discussing the sustainability of Moore's Law. Moore's Law was an observation in the 1960s that the number of transistors on a chip would double every two years. Graphcore's leaders are concerned with a related concept called the Dennard scale, which states that as transistor density increases, power requirements will remain the same.
But this principle is no longer applicable, and now adding more transistors to the chip means that the chip will become hotter and consume more energy. To alleviate this problem, many chip makers design their own products so that they don't exhaust all of their processing power every time, only running the parts necessary to support the application. On the chip, these once unused areas are called "dark silicon."
Knowles and Thun said that unless the circuit can be fundamentally redesigned to increase efficiency, high temperature problems will become a hindrance.Mobile phoneAnd laptops will become faster in the next few years as a major obstacle. Daniel Wilkinson (Daniel Wilkinson), who is in charge of Graphcore chip architecture, said: "I need to start from scratch, which has never happened in chip design."
This can't help but challenge this team of dozens of engineers to design a chip that can take advantage of all the processing power and consume less power than the most advanced GPUs. One of the larger energy pressures of silicon involves moving and retrieving data, but historically, processors and memory are separate. Knowles said that transferring data back and forth between these components is "very energy intensive." Graphcore began to design Knowles' "homogeneous structure", which "mixed" the logic of the chip with the memory so that it didn't need to spend too much energy to transfer the data to other hardware.
Over the past three years, Knowles and Thun have simulated computer test methods for hundreds of chip layouts, and finally determined the design of 1216 processor cores. Knowles called it "a lot of decentralized energy." Processor island." The final IPU debuted in 2018, a very stylish microchip with nearly 24 billion transistors that can access data at a fraction of the power of the GPU. Thun stood in a messy electronics lab at Bristol headquarters, his fingers sliding over the IPU's mirror-like surface: "Each chip has a power of 120 watts, similar to a bright incandescent bulb."
To test the prototype of the chip, the research team provided it with a standard data training model that contained millions of images labeled with common objects (fruit, animals, cars). An engineer then inquired IPU about his own photo of Zeus, and in less than an hour, the computer not only correctly identified it, but also correctly described the appearance of Zeus. Knowles said: "IPU can recognize it as a tabby cat."
Since the first test, the IPU has speeded up and now recognizes more than 10,000 images per second. The goal of the chip is to be able to digest and identify much more complex data models, enabling the system to understand what cats are at a more basic level. Knowles said: "We will not tell the machine what to do, just describe how it should be learned, and give it a lot of examples and data, it does not actually need supervision, the machine is exploring what to do."
Figure 3: Graphcore's first chip Colossus
On the fifth floor of the Graphcore office, cumbersome industryair conditioningBlowing cold air into the company's data server room, swaying the curtains back and forth, let Bristol's unusual sun shine in mid-May. Even though these chips are installed inrefrigeratorOn the size of the box server, it is very energy efficient, but these machines still generate a lot of heat. These IPU server racks are sufficient to perform 64 petaflops, equivalent to 183,000 units.iPhoneX runs at the highest speed at the same time. Knowles and Thun gave their IPU the nickname "Colossus" under the name of the world's first electronically programmable computer. The computer was developed by the British government during the Second World War to crack encrypted information from Germany.
Graphcore has been included from BMW (BMW),Microsoft(Microsoft) and Samsung (Samsung) raised $328 million in funds, and the company's valuation in December was $1.7 billion. Graphcore declined to comment on the specific application of its chip on the grounds of signing a non-disclosure agreement, but considering its investors, many use cases seem obvious, such as self-driving cars, voice assistants like Siri andcloud serviceFarm, etc. But Knowles is most interested in the application of changing human nature. For example, IPU may have a greater impact on the complex analysis that scientists need in climate change and medical research.
To help large corporate customers solve the problem of how to build the next generation of computers to use the chip correctly, Graphcore provides a server blueprint and packages its products with free software tools. Thun said: "We will give you the recipe for computer design, and then sell you the ingredients." IPU relies on the so-called "parallel computing" concept. The basic idea of writing a program is to set the function for each processor, but with the proliferation of on-chip processors (large Graphcore chips include about 5 million processor cores, each time you can run nearly 30 million programs), this The coding task has replaced the manual programming, which means that the processor must be programmed automatically to execute independently.
In layman's terms, Graphcore divides huge computing tasks into small data problems, each of which is handled separately on these "processor islands" and then synchronized like the Marine Corps military band, at the most efficient Always share what they have learned.
Tobias Yang (Tobias Jahn), chief investor in BMW's venture capital division, envisions the use of Graphcore chips in the company's data centers, perhaps including its cars. "BMW intends to make Graphcore a large global silicon supplier," he said. Autopilot cars have to perform too many critical tasks immediately, making them a key market for products such as IPU, as working in cloud computing tends to delay. Arm Holdings co-founder Hauser (Hauser) estimates that each driverless car may need two IPU.Graphcore, saying its revenue is expected to reach $50 million in 2019.
Big-name competitors have also poured into this field. Electric car manufacturer Tesla recently patented its own AI chip, and Google last year introduced a microprocessor designed specifically for machine learning. NVIDIA has been improving its main GPU chip design to make it more inaccurate but more efficient, which is more like Graphcore's approach.
Alan Priestly (Alan Priestley), an analyst at Gartner, a market research firm, said: "all the other companies are knocking on Nvidia's door. Graphcore has a big advantage, but it is still a very small competitor compared to Nvidia's market share. So, while their IPU may outperform Nvidia's GPU, in these workload, the risk is that customers tend to choose 'good enough' rather than 'excellence'. "
If, as promised, IPU can make machines run 100 times more powerful than today's computers, another major challenge will be ethical dilemmas. Thun and Knowles are wary of these dangers, especially how these technologies can be abused by weapons and surveillance. However, they said that the government will eventually need to set limits. Knowles pointed out: "Mechanical power helped us invent airplanes and cars, but it also helped invent the tank. Over time, society will have to find a balance between good and evil."
At present, Graphcore is focused on developing more software, allowing customers to see the power of IPU while expanding the business to the final stage of the market. For every major milestone, the company celebrates with a bottle of champagne, such as $50 million in financing at the end of 2017 and $10 million in sales orders in 2018. This sign of growth is everywhere in Graphcore's office, and the champagne bottles are getting bigger and bigger.
Knowleson always starts with the Pol Roger Champagne brand, which they believe represents their pride and they may help the UK to be the first tech giant. Knowles said: "Starting with Pol Roger and ending with Pol Roger." Knowles recently drank a 9-litre bottle of champagne. He said: "When you make an initial public offering (IPO), you will open the largest bottle. Champagne."