The work, shown at the automaker’s live-streamed AI Day, demonstrates the company’s continuing pursuit of vertical integration.

The D1 chip, part of Tesla’s Dojo supercomputer system, uses a 7-nanometer manufacturing process, with 362 teraflops of processing power, said Ganesh Venkataramanan, senior director of Autopilot hardware. Tesla places 25 of these chips onto a single “training tile,” and 120 of these tiles come together across several server cabinets, amounting to over an exaflop of power, Venkataramanan said:

“We are assembling our first cabinets pretty soon,” said Venkataramanan, who previously worked at chipmaker AMD.

He said the Tesla technology will be the fastest AI-training computer. Chipmaker Intel, graphics card maker Nvidia and start-up Graphcore are among the companies that make chips that companies can use to train AI models. The chips can help train models for recognizing a variety of items from video feeds collected by cameras inside Tesla vehicles. Model training requires extensive computing work. CEO Elon Musk said

“We should have Dojo operational next year.”

The work comes two years after Tesla began producing vehicles containing AI chips it built in house. Those chips help the car’s onboard software make decisions very quickly in response to what’s happening on the road.

Tesla currently offers what it calls a “Full Self-Driving Capability” add-on for new vehicles. The $10,000 package lets the car automatically change lanes, navigate on highways, move into parking spots and emerge from a parking spot to arrive by the driver. The Tesla website says later this year the package will also include the ability for a Tesla to automatically steer on city streets, although Tesla had previously promised that feature would come out in 2019.

Tesla’s website says Full-Self Driving elements “require active driver supervision and do not make the vehicle autonomous.” Earlier this year Tesla introduced a $199 monthly subscription for Tesla owners who wish to access Full-Self Driving.

Critics have said that Tesla’s moniker for its driver-assistance features is misleading. Tesla’s software does not offer Level 5 autonomy, where a car can completely drive in all circumstances without human intervention. Last year, a German court ruled that Tesla had misled consumers about the autonomous capabilities of its cars. The National Highway Traffic Safety Administration announced an investigation of Tesla’s Autopilot automatic steering, accelerating and braking capability earlier this week, following a string of crashes.

One of the flipsides of our runaway internet usage is its impact on climate due to the massive amount of electricity consumed by computer servers. Current CO2 emissions from data centres are as high as from global air traffic combined—with emissions expected to double within just a few years.

Only a handful of years have passed since Professor Mikkel Thorup was among a group of researchers behind an algorithm that addressed part of this problem by producing a groundbreaking recipe to streamline computer server workflows. Their work saved energy and resources. Tech giants including Vimeo and Google enthusiastically implemented the algorithm in their systems, with online video platform Vimeo reporting that the algorithm had reduced their bandwidth usage by a factor of eight.

Now, Thorup and two fellow UCPH researchers have perfected the already clever algorithm, making it possible to address a fundamental problem in computer systems—the fact that some servers become overloaded while other servers have capacity left—many times faster than today. Professor Thorup of the University of Copenhagen’s Department of Computer Science, who developed the algorithm alongside department colleagues Anders Armand and Jakob Bæk Tejs Knudsen, commented:

“We have found an algorithm that removes one of the major causes of overloaded servers once and for all. Our initial algorithm was a huge improvement over the way the industry had been doing things, but this version is many times better and reduces resource usage to the greatest extent possible. Furthermore, it is free to use for all.”

The algorithm addresses the problem of servers becoming overloaded as they receive more requests from clients than they have the capacity to handle. This happens as users pile in to watch a certain Vimeo video or Netflix film. As a result, systems often need to shift clients around many times to achieve a balanced distribution among servers.

The mathematical calculation required to achieve this balancing act is extraordinarily difficult as up to a billion servers can be involved in the system. And, it is ever-volatile as new clients and servers join and leave. This leads to congestion and server breakdowns, as well as resource consumption that influences the overall climate impact. Thorup also explains

“As internet traffic soars explosively, the problem will continue to grow. Therefore, we need a scalable solution that doesn’t depend on the number of servers involved. Our algorithm provides exactly such a solution.”

According to the American IT firm Cisco, internet traffic is projected to triple between 2017 and 2022. Next year, online videos will make up 82% of all internet traffic. The new algorithm ensures that clients are distributed as evenly as possible among servers, by moving them around as little as possible, and by retrieving content as locally as possible.

For example, to ensure that client distribution among servers balances so that no server is more than 10% more burdened than others, the old algorithm might deal with an update by moving a client one hundred times. The new algorithm reduces this to 10 moves, even when there are billions of clients and servers in the system. Mathematically stated: If the balance is to be kept within a factor of 1+1/X, the improvement in the number of moves from X2 to X is generally impossible to improve upon.

As many large IT firms have already implemented Professor Thorup’s original algorithm, he believes that the industry will adopt the new one immediately—and that it may already be in use. Studies have demonstrated that global data centres consume more than 400 terawatt-hours of electricity annually. This accounts for approximately 2% of the world’s total greenhouse gas emissions and currently equals all emissions from global air traffic. Datacenter electricity consumption is expected to double by 2025.

According to the Danish Council on Climate Change, a single large data centre consumes the equivalent of 4% of Denmark’s total electricity consumption.

Mikkel Thorup is head of the BARC research centre (Basic Algorithms Research Copenhagen) at the University of Copenhagen’s Department of Computer Science. BARC has positioned Copenhagen as the world’s fourth-best place in basic research in the design and analysis of algorithms. BARC is funded by the VILLUM FOUNDATION.

I admit I hadn’t considered software design as a major contributor to greenhouse gas. But the folks at the foundation make a compelling case that as everything becomes more connected, the way we build the software can have profound impacts. Asim Hussain, green cloud advocacy lead at Microsoft and executive director of the Green Software Foundation, commented in an email:

“We often look at sustainability in tech through the lens of hardware, focusing on how to make hardware more carbon-efficient. Relatively little has been done to look at sustainability through the lens of software. From where we are sitting, there seem to be some ripe easy pickings for sustainability improvements by looking at how we run software on hardware.”

What is ‘green software’?

Green software takes into account sustainability considerations and climate science to design software that is “carbon-efficient” — meaning it minimizes the number of carbon emissions per unit of work. It takes into account factors such as energy efficiency within its design.

At the risk of getting wonky, the foundation broadly divides this software into different buckets:

• Carbon-efficient applications that more efficiently use resources linked to carbon emissions.
• Carbon-aware applications that change behaviour to minimize carbon emissions.

Cumulatively, this type of software could have a significant impact. The software runs billions of internet-connected devices, from phones and refrigerators to data centres and personal computers. The foundation says that 20% of electricity output will be consumed by information and communications technology by 2030. The group has a goal to reduce emissions related to the industry’s software by 45% by 2030.

The foundation’s strategy is three-pronged: develop standards and best practices to make software consume less energy; support open-source innovation; and facilitate an international community. Hussain added:

“We think we can pull some pretty big levers through just growing the ecosystem of people with the skills to build software that emits less carbon, as well as creating standards that organizations can easily adopt or that might influence policy decisions.”

The data centre example

Data centres are energy hogs. They’re also essential to our ever-connected world, and their energy demands are huge and expected to grow. In 2018, research suggests data centres were responsible for about the same amount of carbon emissions as aviation. Today, they’re responsible for 1% of global electricity demand, with projections that the segment will consume 3 to 8% within the next decade.

While that is a tremendous amount of energy, data centre consumption is less than we thought it would be 15 years ago. Energy use in data centres doubled from 2000 to 2005, a trend we thought would continue. Between 2010 and 2018, computing output jumped sixfold — but energy consumption rose only 6%.

In part, that’s thanks to the rise of cloud computing, which allows companies and individuals to use centralized data centres for computing and storage. Additionally, data centres have become more efficient. The software has had a big role to play and can do much more. Google, for example, is working towards 24/7 clean energy for its data centres with the help of carbon-aware software.

These increased efficiencies illustrate how difficult it is to predict the future of disruptive technologies. It also reflects how tech companies, founded on the ethos of innovation, are positioned to reinvent their processes.

Nvidia is talking up AI at Computex 2021, which is returning to Taiwan as a hybrid in-person and virtual event, that will be held tomorrow. Nvidia and the National Energy Research Scientific Computing Center (NERSC) last week flipped the “on” switch for Perlmutter, billed as the world’s fastest supercomputer for AI workloads.

The new Base Command platform gives developers access to the cloud-hosted computing power of Nvidia DGX SuperPOD AI supercomputers and NetApp data management tools.

Nvidia touted Base Command as a way for enterprise developers to “quickly move their AI projects from prototypes to production” with software designed “for large-scale, multi-user and multi-team AI development workflows hosted either on-premises or in the cloud,” the company said. The software development platform enables numerous researchers and data scientists to “simultaneously work on accelerated computing resources, helping enterprises maximize the productivity of both their expert developers and their valuable AI infrastructure,” the company said.

Base Command offers a single pane of glass interface to view AI software development on integrated monitoring and reporting dashboards, with command-line APIs available.

Nvidia also announced the availability of new AI-optimized servers from major computer manufacturers as part of its Nvidia-certified systems program. The new systems are certified to run Nvidia AI Enterprise software and are either available now or coming later this year from OEMs. Participating companies include Advantech, Altos, ASRock Rack, Asus, Dell Technologies, Gigabyte, Hewlett Packard Enterprise, Lenovo, Nettrix, QCT, and Supermicro.

New x86 servers based on Nvidia Ampere architecture GPUs are available now, the company said. Nvidia-certified systems using Bluefield-2 DPUs will come out later this year and non-x86 machines powered by Arm CPUs will arrive in 2022. Manuvir Das, Nvidia head of enterprise computing, said:

“Enterprises across every industry need to support their innovative work in AI on traditional data centre infrastructure. The open, growing ecosystem of Nvidia-certified systems provides unprecedented customer choice in servers validated by Nvidia to power world-class AI.”

Das said the new servers will become “some of the highest-volume x86 servers used in mainstream data centres, bringing the power of AI to a wide range of industries, including health care, manufacturing, retail, and financial services.”

The certified systems are expected to run software such as the Nvidia AI Enterprise suite of AI and data analytics software on VMware vSphere, Nvidia Omniverse Enterprise for design collaboration and advanced simulation, and Red Hat OpenShift for AI development, with additional support for Cloudera data engineering and machine learning modelling tools.