Machine learning via cloud for semiconductor industry

Estimated reading time: 5 minutes

During SEMICON SEA last week, Google’s Head of Customer Engineering Singapore and Malaysia, Google APAC, shared about some of the interesting work his organisation is doing to enable artificial intelligence and machine learning (AI and ML) use cases for the semiconductor industry.

These AI and ML deployments are relevant for a wide range of industries, not just semiconductors.

As such, hyperscalers are scrambling to bring the value of AI and ML to business, organisations, and many more industries, and this is one of many drivers for the many data centre expansion plans, to date.But, AI and ML deployments themselves require massive compute power in the back end.

From the semiconductor industry point of view, we still need a lot of chip designs, a lot of chip manufacturing to fuel the economy.

“However some fundamentals remain unchanged, and if anything they have become only more challenging. Critical businesses and business processes must remain open. From the semiconductor industry point of view, we still need a lot of chip designs, a lot of chip manufacturing to fuel the economy.”

But won’t the current way of working with remote work and virtual collaboration tools, impact chip designing and manufacturing?

Automation and digitalisation have not halted during this Not Normal times. In fact, they have accelerated and the foundational technologies that power them, has to keep up if not progress even faster.

Karthick said, “At Google, we are helping our customers achieve the next level of efficiency, with or without humans.”

Cloud building experience speaks

Google boasts nine key products that have over a billion users each. “And each of these products depend on the Google Cloud infrastructure.

Recognising this, Google has specific hardware like Google TPUs (tensor processing units) that allow the adoption of AI across many new use cases / devices / architectures. Karthik Ramarao, Google’s head of customer engineering, pointed out how technologies have enabled businesses to continue their operations, albeit with new ways of working.

“It goes without saying that Google Cloud is their priority number one, and its one of the fastest growing businesses for them. “Our core business depends on the heart of Google Cloud and that is the security that been built ground up, and one of the largest privately managed networks,” Karthik said.

Huge investments have gone into custom-built hardware and infrastructure, so that Google is able to deliver, end to end security solutions from data centre, all the way to the edge devices.

Another way Google sets itself apart from most of its competitors is their belief in open cloud. “Our customers should be able to pick the most valuable and the most suitable cloud for their strategy.

This includes customers having the option to leverage their previous investments in on-premise infrastructure, and to burst into the cloud for specific workloads as and when they need to.

Karthik explained, “They should have choice of a multi-cloud model, where they can choose their cloud service provider for each of their applications, without a hurdle in moving data across, or recode anything in order to move from cloud to cloud.”

Besides making the cloud scalable, one notable thing they have embedded into everything is intelligence.

He explained, “We make it easy for the customers to apply AI and ML in their own businesses to automatically do things in the semiconductor industry, for example, like wafer level defect detection, or predictive maintenance, or fab process optimisation.”

AI and ML help to answer questions around how to use achieve specific metrics and parameters for the semiconductor industry.

Security at all layers

Google designed hardware which they believe offers security across every layer of the Google Cloud.”We designed the Titan chip to establish hardware root of trust. We purpose-built everything from servers to storage, to network to data centres, and hence we establish a chain of trust across every single component and every single element in our cloud.

Workload data stays within Google networks, within their own trusted environments.

“This, as well as a automatic encryption, ensures a highly secure data sharing environment between chip design teams, verification teams, across the globe with remote foundries and other ecosystem partners.”We do this (data privacy and data security) for ourselves and are able to extend this to our customers.”

Google lays claim to being only service provider to own its own private fibre cables and subsea cables around the world. “We invested about USD10 billion, and we continue to extend it for our customers so they can manage their workload from anywhere, as though they are on one virtual private cloud.”

Karthik also said that workload data stays within Google networks, within their own trusted environments. “It takes the shortest path and never leaves our network therefore giving our customers the lowest latency and the highest level of security.”

Compute-intensive machine learning

Electronic Design Automation, or EDA, which is at the core of chip design, needs to use a lot of compute power which Google Cloud can fulfill. It expects to be able to provide the same performance as an on-premise infrastructure, with the notable difference of allowing customers to scale up or scale down the usage of resources.But there are further advantages.

Karthik added, “Imagine that you can connect your chip design floor to the latest and the greatest AI technology.” These are custom-built AI accelerator chips, or TPUs, that were built to accelerate machine learning workloads.

Exaflop power usually reserved for supercomputers are now available for access via Google Cloud. So, chip designers can pass an on-prem computational workload to the Google cloud and have it be completed in a fraction of the time.

It will take hours instead of weeks or months to train ML models with the help of powerful TPUs.

Google’s AI and ML mission

According to Karthik, Google’s mission is to democratise AI and ML.

Beginners can have a taste of AI and ML, while enthusiasts can train their own data sets with auto-ML. Organisations that have their own data scientists may even use Google’s ready end-to-end AI resource, to build models to train and deploy.

Karthik concluded, “It is much more than just compute, and just services around compute. It is about providing an end to end, ability for an organisation to adopt AI and ML, in order to get the best out of their own infrastructure.”

For the semiconductor industry, that means operational efficiency and better yield for the chip design process as well as the chip manufacturing process.