Today, ML-driven innovation is fundamentally transforming computing, enabling entirely new classes of internet services. For example, recent state-of-the-art lage models such as PaLM and Chinchilla herald a coming paradigm shift where ML services will augment human creativity. All indications are that we are still in the early stages of what will be the next qualitative step function in computing. Realizing this transformation will require democratized and affordable access through cloud computing where the best of compute, networking, storage, and ML can be brought to bear seamlessly on ever larger-scale problem domains.

Today’s release of MLPerf™ 2.0 results from the MLCommons® Association highlights the public availability of the most powerful and efficient ML infrastructure anywhere. Google’s TPU v4 ML supercomputers set performance records on five benchmarks, with an average speedup of 1.42x over the next fastest non-Google submission, and 1.5x vs our MLPerf 1.0 submission. Even more compelling — four of these record runs were conducted on the publicly available Google Cloud ML hub that we announced at Google I/O. ML Hub runs out of our Oklahoma data center, which uses over 90% carbon-free energy.

Let’s take a closer look at the results.

Performance at scale…and in the public cloud

Our 2.0 submissions1, all running on TensorFlow, demonstrated leading performance across all five benchmarks. We scaled two of our submissions to run on full TPU v4 Pods. Each Cloud TPU v4 Pod consists of 4096 chips connected together via an ultra-fast interconnect network with an industry-leading 6 terabits per second (Tbps) of bandwidth per host, enabling rapid training for the largest models.

Hardware aside, these benchmark results were made possible in no small part by our work to improve the TPU software stack. Scalability and performance optimizations in the TPU compiler and runtime, including faster embedding lookups and improved model weight distribution across the TPU pod, enabled much of these improvements, and are now widely available to TPU users. For example, we made a number of performance improvements to the virtualization stack to fully utilize the compute power of both CPU hosts and TPU chips to achieve peak performance on image and recommendation models. These optimizations reflect lessons from Google’s cutting-edge internal ML use cases across Search, YouTube, and more. We are excited to bring the benefits of this work to all Google Cloud users as well.

Translating MLPerf wins to customer wins

Cloud TPU’s industry-leading performance at scale also translates to cost savings for customers. Based on our analysis summarized in Figure 3, Cloud TPUs on Google Cloud provide ~35-50% savings vs A100 on Microsoft Azure (see Figure 3). We employed the following methodology to calculate this result:2

We compared the end-to-end times of the largest-scale MLPerf submissions, namely ResNet and BERT, from Google and NVIDIA. These submissions make use of a similar number of chips — upwards of 4000 TPU and GPU chips. Since performance does not scale linearly with chip count, we compared two submissions with roughly the same number of chips.

To simplify the 4216-chip A100 comparison for ResNet vs our 4096-chip TPU submission, we made an assumption in favor of GPUs that 4096 A100 chips would deliver the same performance as 4216 chips.

For pricing, we compared our publicly available Cloud TPU v4 on-demand prices ($3.22 per chip-hour) to Azure’s on-demand prices for A1003 ($4.1 per chip-hour). This once again favors the A100s since we assume zero virtualization overhead in moving from on-prem (NVIDIA’s results) to Azure Cloud.

The savings are especially meaningful given that real-world models such as GPT-3 and PaLM are much larger than the BERT and ResNet models used in the MLPerf benchmark: PaLM is a 540 billion parameter model, while the BERT model used in the MLPerf benchmark has only 340 million parameters — a 1000x difference in scale. Based on our experience, the benefits of TPUs will grow significantly with scale and make the case all the more compelling for training on Cloud TPU v4.

Have your cake and eat it too — a continued focus on sustainability

Performance at scale must take environmental concerns as a primary constraint and optimization target. The Cloud TPU v4 pods powering our MLPerf results run with 90% carbon-free energy and a Power Usage Efficiency of 1.10, meaning that less than 10% of the power delivered to the data center is lost through conversion, heat, or other sources of inefficiency. The TPU v4 chip delivers 3x the peak FLOPs per watt relative to the v3 generation. This combination of carbon-free energy and extraordinary power delivery and computation efficiency makes Cloud TPUs among the most efficient in the world.4

Making the switch to Cloud TPUs

There has never been a better time for customers to adopt Cloud TPUs. Significant performance and cost savings at scale as well as a deep-rooted focus on sustainability are why customers such as Cohere, LG AI Research, Innersight Labs, and Allen Institute have made the switch. If you are ready to begin using Cloud TPUs for your workloads, please fill out this form. We are excited to partner with ML practitioners around the world to further accelerate the incredible rate of ML breakthroughs and innovation with Google Cloud’s TPU offerings.

1. MLPerf™ v2.0 Training Closed. Retrieved from https://mlcommons.org/en/training-normal-20/ 29 June 2022, results 2.0-2010, 2.0-2012, 2.0-2098, 2.0-2099, 2.0-2103, 2.0-2106, 2.0-2107, 2.0-2120. The MLPerf name and logo are trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information.
2. MLPerf v1.0 and v2.0 Training Closed. Retrieved from https://mlcommons.org/en/training-normal-20/ 29 June 2022, results 1.0-1088, 1.0-1090, 1.0-1092, 2.0-2010, 2.0-2012, 2.0-2120.
3. ND96amsr A100 v4 Azure VMs, powered by eight 80 GB NVIDIA Ampere A100 GPUs (Azure’s flagship Deep Learning and Tightly Coupled HPC GPU offering with CentOS or Ubuntu Linux) is used for this benchmarking
4. Cost to train is not an official MLPerf metric and is not verified by MLCommons Association. Azure performance is a favorable estimate as described in the text, not an MLPerf result. Computations are based on results from MLPerf v2.0 Training Closed. Retrieved from https://mlcommons.org/en/training-normal-20/ 29 June 2022, results 2.0-2012, 2.0-2106, 2.0-2107.

Background

At IKEA we have multiple places in our customer journey in various channels where different kinds of personalization can deliver a superior customer experience. Product recommendations in the shopping basket, content recommendations in editorial sections, inspirational recommendations on product pages and more. After a while in the broader “recommendations” team there was a decision to split the team to have one sub-team focused on product recommendations. The pandemic altered customer behavior and needs as well. At that inflection point we decided to change our way of working and dive head-first into a more scientific approach to handle the operational complexities of delivering high quality product recommendations at scale. We deemed this necessary to improve our level of personalization and to have a holistic understanding of our customers.

Data Driven Decisions

The first step was to radically improve our ability to get high-quality quantitative information to understand how our ‘recommendation’ solutions affected personalization. We did this through high volume A/B testing on customer behaviour and after initial experimentation, we had a few key learnings:  

1. The mix of both UX and algorithms are really important for a cohesive customer experience. 
2. The quality of personalization can’t be measured in silos. Statistical significance  can be attained  by testing several groups of recommendations at once.

Once we came up with a solid framework for gathering data and acknowledged how little we knew about our customers, we were able to explore an incredible number of creative options – nothing was off the table. This was a very humbling experience, in that it opened up new perspectives for personalization, a more curious and less confined way of thinking. We learned to trust the data because it might show you things you don’t expect. 

Experimentation and Learning Framework

Our teams created ways to quickly deploy experimental modifications to our existing solution. This enabled experimentation in the front-end with the user experience, including details in headings and images. This also covered tweaks in the backend with anything from detailed manual additions or removals of recommendations to mixing and matching of various algorithms both home grown and from Recommendations AI.

This flexibility came with an overhead–more complexity and cost relative to directly retrieving recommendations from Recommendations AI. However, the benefit was that we were no longer dependent on manual evaluation of what made for a good recommendation system. We aligned on a data-driven and qualitative approach to provisioning recommendations and significantly accelerated our experimentation timeline. Together with optimization of the CI/CD pipeline this enabled the team to take an idea or hypothesis from inception to A/B testing with customers in less than half an hour.

Recommendations AI Experiments

Our team’s infrastructure was already running on GCP and when we received early access to Recommendations AI, the requirements to get started were minimal and that allowed us to start with initial tests requiring minimal effort and investment.

We started with a few use-cases and identified places where our existing recommendation algorithms needed improvement or complementary recommendations. We also explored additional ways where more useful information could be presented to the customers through personalized recommendations. 

Recommendations AI Model Combinations

While Recommendations AI might be considered a simple API to get a set of product recommendations, as we dove deeper into the solution it became apparent that it could be tweaked in several different ways to offer many fine tuning configurations to meet business goals. While too much fine tuning and customization could lead to subpar performance, in general we found that it was a great strategy to give us several versions of ML powered recommendations to work with. The further you personalize the experience, the more options you have to likely pick the best one for the customer.

Recommendations AI models like  ‘Recommended for you’, ‘Frequently Bought Together’ and ‘Others you may like’; are coupled with business goals like optimizing for conversion rate, click through rate and revenue. We experimented with many different model combinations and custom rules. All this was easily configurable right in the GCP console. One of the simplest custom configurations we used was to only recommend items that were in stock, and when items were out of stock we looked at similar items that were available to augment the experience. 

Collaboration with Google

Our collaboration with Google Cloud accelerated our learning process during experimentation. We worked closely together early in the product development. Additionally, their model provided flexibility to change direction and allow for more options than we had previously. Ultimately, this provided us a way to drastically improve our time to market with a product that produced tremendous results that we could not have accomplished on our own.

Results and Takeaways

With more personalized and real-time recommendations available we saw great success. We were able to increase the number of relevant recommendations displayed on a page by +400%. To accommodate the wider repertoire of recommendations we had to change the user experience. For example, in some places we had horizontally scrolling displays of product recommendations which were much easier for customers to use.

Another consequence of displaying more personalized recommendations was tangible improvement to conversion rate and average order value. Recommendations AI algorithms helped customers in two ways: 

1. Customers were able to find products that they liked quickly and establish their preferred choice among other options more quickly as well, giving them confidence to make a purchase through much fewer clicks. Even though we previously already had well tuned recommendations of several types, with Recommendations AI we measured +30% improvement in click through rates. 
2. Average order value saw a +2% surge with numerous examples of how Recommendations AI could help customers find both attractive and directly complementary products, expanding the customer purchase from a single product to an entire home furnishing solution.

As a direct effect of having stronger business results, the team started exploring more places in the customer journey where our growing buffet of recommendations could be used. We’d start with an initial experiment to answer if displaying recommendations in the specific context made sense at all. Frequently the data that emerged from these experiments prodded us to iterate further on what additional types of recommendations would be most appropriate to show to the customer as the customer’s behaviour evolved. Today, most of IKEA’s site recommendations are powered by Recommendations AI.

One key takeaway is that for some types of personalized recommendations there are benefits to using advanced algorithms that require a lot of high level data science and engineering competence to build since they outperform simplistic approaches. In some places, simplistic approaches work very well and in others the right decision is to not have product recommendations at all. For an effective use of product recommendations you need to have all the above options and the ability to tell when to use which one.

Next steps

When working with something so tightly related to customer experience, there is a constant change in user behaviour and new learnings to observe and adapt to. Product recommendations are rarely the main stand alone experience and frequently something that is used to help and enhance an experience. We see a lot of value in having a large toolbox of possible options and a team with a relentless focus on collaboration to improve the customer experience. We’re working directly with the Recommendations AI team and experimenting with several new features that we’re excited about. 

In the future we see opportunities of improving the customer journey through a more visual experience that inspires the customer rather than relying on customers to use their imagination to visualize groups of products together. Vision Product Search provides that and is something we’re looking into deploying next. We’ll be sharing more about our journey with Recommendations AI at the Google Cloud Retail Summit session ‘IKEA’s Approach to Building a Powerful Recommendations Engine’ on July 27th 2021.

Best wishes to all developers from the IKEA product recommendations team & the Google Recommendations AI team!

Expectations for both consumer and commercial products have changed. Consumers want products that evolve with their needs, adapt to their preferences, and stay up-to-date over time. Manufacturers, in turn, need to create products that provide engaging customer experiences not only to better compete in the marketplace, but also to provide new monetization opportunities. 

However, embedding intelligence into new and existing products is challenging. Updating hardware is costly, and existing connected products do not have the capability to add new features. Furthermore, manufacturers do not have sufficient customer insights due to product telemetry and customer data silos, and may lack the AI expertise to quickly develop and deploy these features. 

That’s why today we’re launching Intelligent Products Essentials, a solution that allows manufacturers to rapidly deliver products that adapt to their owners, update features over-the-air using AI at the edge, and provide customer insights using analytics in the cloud. The solution is designed to assist manufacturers in their product development journeys—whether developing a new product or enhancing existing ones. 

With Intelligent Products Essentials, manufacturers can:

• Personalize customer experiences: Provide a compelling ownership experience that evolves over the lifetime of the product. For example, a chatbot that contextualizes responses based on product status and customer profile.
• Manage and update products over-the-air: Deploy updates to products in the field,  gather performance insights and evolve capabilities over time with monetization opportunities.
• Predict parts and service issues: Detect operating thresholds, anomalies and predict failures to proactively recommend service using AI, reducing warranty claims, decreasing parts shortages and increasing customer satisfaction.

In order to help manufacturers quickly deploy these use cases and many more, Intelligent Products Essentials provides the following:

• Edge connections: Connect and ingest raw or time-series product telemetry from various device platforms utilizing IoT Core or Pub/Sub and enable deployment and management of firmware over-the-air and machine learning models with Vertex AI at the edge.
• Ownership App Template: Easily build connected product companion apps that work on smartphones, tablets, and computers. Use a pre-built API and accompanying sample app that can incorporate product or device registration, identity management, and provide application behavior analytics using Firebase.
• Product fleet management: Manage, update and analyze fleets of connected products via APIs, Google Kubernetes Engine, and Looker.
• AI services: Create new features or capabilities for your products using AI and machine learning products such as DialogFlow, Vision AI, AutoML, all from Vertex AI.

Enterprise data integration: Integrate data sources such as Enterprise Asset Management (EAM), Enterprise Resource Planning (ERP), Customer Relationship Management (CRM) systems and others using Dataflow and BigQuery.

Intelligent Products Essentials helps manufacturers build new features across consumer, industrial, enterprise, and transportation products. Manufacturers can implement the solution in-house, or work with one of our certified solution integration partners like Quantifi and Softserve. 

“The focus on intelligent products that Google Cloud is deploying provides a digital option for manufacturers and users. At its heart, systems like Intelligent Product Essentials are all about decision making. IDC sees faster and more effective decision-making as the fundamental reason for the drive to digitize products and processes. It’s how you can make faster and more effective decisions to meet heightened customer expectations, generate faster cash flow, and better revenue realization,” said Kevin Prouty, Group Vice President at IDC. “Digital offerings like Google’s Intelligent Product Essentials potentially go the last mile with the ability to connect the digital thread all the way through to the final user.”

Customers adopting Intelligent Products Essentials

GE Appliances, a Haier company, are enhancing their appliances using new AI-powered intelligent features to enable:

• Intelligent cooking: Help cook the perfect meal to personal preferences, regardless of your expertise and abilities in the kitchen.
• Frictionless service: Build smart appliances that know when they need maintenance and make it simple to take action or schedule services.
• Integrated digital lifestyle: Make appliances useful at every step of the way by integrating them with digital lifestyle services – for example, automating appliance behaviors according to customer calendars, such as oven preheating or scheduling the dishwasher to run in the late evening.

“Intelligent Products Essentials enhances our smart appliances ecosystem, offering richer consumer habit insights. This enables us to develop and offer new features and experiences to integrate with their digital lifestyle.“ —Shawn Stover, Vice-president Smart Home Solutions at GE Appliances.https://www.youtube.com/embed/zaWAJN8aKOw?enablejsapi=1&

Serial 1, Powered by Harley-Davidson, is using Intelligent Product Essentials to manage and update its next generation eBicycles, and personalize its customers’ digital ownership experiences. 

“At Serial 1, we are dedicated to creating the easiest and most intuitive way to experience the fun, freedom, and adventure of riding a pedal-assist electric bicycle. Connectivity is a key component of delivering that mission, and working together to integrate Intelligent Product Essentials into our eBicycles will ensure that our customers enjoy the best possible user experience.”— Jason Huntsman, President, Serial 1. 

Magic Leap, an augmented reality pioneer with industry-leading hardware and software, is building field service solutions with Intelligent Products Essentials with the goal of connecting manufacturers, dealers, and customers to more proactive and intelligent service.

“We look forward to using Intelligent Products Essentials to enable us to rapidly integrate manufacturers’ product data with dealer service partners into our field service solution. We’re excited to partner with Google Cloud as we continue to push the boundaries of physical interaction with the digital world.” — Walter Delph, Chief Business Officer, Magic Leap

Intelligent Product Essentials is available today. To learn more, visit our website.