The latest round of MLPerf benchmark results have been released, and Google’s TPU v4 supercomputers demonstrated record-breaking performance at scale. This is a timely milestone since large-scale machine learning training has enabled many of the recent breakthroughs in AI, with the latest models encompassing billions or even trillions of parameters (T5, Meena, GShard, Switch Transformer, and GPT-3). 

Google’s TPU v4 Pod was designed, in part, to meet these expansive training needs, and TPU v4 Pods set performance records in four of the six MLPerf benchmarks Google entered using TensorFlow and JAX. These scores are a significant improvement over our winning submission from last year and demonstrate that Google once again has the world’s fastest machine learning supercomputers. These TPU v4 Pods are already widely deployed throughout Google data centers for our internal machine learning workloads and will be available via Google Cloud later this year.

Figure 1: Speedup of Google’s best MLPerf Training v1.0 TPU v4 submission over the fastest non-Google submission in any availability category – in this case, all baseline submissions came from NVIDIA. Comparisons are normalized by overall training time regardless of system size. Taller bars are better.¹

Let’s take a closer look at some of the innovations that delivered these ground-breaking results and what this means for large model training at Google and beyond. 

Google’s continued performance leadership

Google’s submissions for the most recent MLPerf demonstrated leading top-line performance (fastest time to reach target quality), setting new performance records in four benchmarks. We achieved this by scaling up to 3,456 of our next-gen TPU v4 ASICs with hundreds of CPU hosts for the multiple benchmarks. We achieved an average of 1.7x improvement in our top-line submissions compared to last year’s results. This means we can now train some of the most common machine learning models in a matter of seconds.

Figure 2: Speedup of Google’s MLPerf Training v1.0 TPU v4 submission over Google’s MLPerf Training v0.7 TPU v3 submission (exception: DLRM results in MLPerf v0.7 were obtained using TPU v4). Comparisons are normalized by overall training time regardless of system size. Taller bars are better. Unet3D not shown since it is a new benchmark for MLPerf v1.0.²

We achieved these performance improvements through continued investment in both our hardware and software stacks. Part of the speedup comes from using Google’s fourth-generation TPU ASIC, which offers a significant boost in raw processing power over the previous generation, TPU v3. 4,096 of these TPU v4 chips are networked together to create a TPU v4 Pod, with each pod delivering 1.1 exaflop/s of peak performance.

Figure 3: A visual representation of 1 exaflop/s of computing power. If 10 million laptops were running simultaneously, then all that computing power would almost match the computing power of 1 exaflop/s.

In parallel, we introduced a number of new features into the XLA compiler to improve the performance of any ML model running on TPU v4. One of these features provides the ability to operate two (or potentially more) TPU cores as a single logical device using a shared uniform memory access system. This memory space unification allows the cores to easily share input and output data – allowing for a more performant allocation of work across cores. A second feature improves performance through a fine-grained overlap of compute and communication. Finally, we introduced a technique to automatically transform convolution operations such that space dimensions are converted into additional batch dimensions. This technique improves performance at the low batch sizes that are common at very large scales.

Enabling large model research using carbon-free energy

Though the margin of difference in topline MLPerf benchmarks can be measured in mere seconds, this can translate to many days worth of training time on the state-of-the-art models that comprise billions or trillions of parameters. To give an example, today we can train a 4 trillion parameter dense Transformer with GSPMD on 2048 TPU cores. For context, this is over 20 times larger than the GPT-3 model published by OpenAI last year. We are already using TPU v4 Pods extensively within Google to develop research breakthroughs such as MUM and LaMDA, and improve our core products such as Search, Assistant and Translate. The faster training times from TPUs result in efficiency savings and improved research and development velocity. Many of these TPU v4 Pods will be operating at or near 90% carbon free energy. Furthermore, cloud datacenters can be ~1.4-2X more energy efficient than typical datacenters, and the ML-oriented accelerators – like TPUs – running inside them can be ~2-5X more effective than off-the-shelf systems. 

We are also excited to soon offer TPU v4 Pods on Google Cloud, making the world’s fastest machine learning training supercomputers available to customers around the world. Cloud TPUs support leading frameworks such as TensorFlow, PyTorch, and Jax, and we recently released an all-new Cloud TPU system architecture that provides direct access to TPU host machines, greatly improving the user experience. 

Want to learn more? 

Please contact your Google Cloud sales representative to request early access to Cloud TPU v4 Pods. We are excited to see how you will expand the machine learning frontier with access to exaflops of TPU computing power!

^{1. All results retrieved from www.mlperf.org on June 30, 2021. MLPerf name and logo are trademarks. See www.mlperf.org for more information. Chart uses results 1.0-1067, 1.0-1070, 1.0-1071, 1.0-1072, 1.0-1073, 1.0-1074, 1.0-1075, 1.0-1076, 1.0-1077, 1.0-1088, 1.0-1089, 1.0-1090, 1.0-1091, 1.0-1092.
2. All results retrieved from www.mlperf.org on June 30, 2021. MLPerf name and logo are trademarks. See www.mlperf.org for more information. Chart uses results 0.7-65, 0.7-66, 0.7-67, 1.0-1088, 1.0-1090, 1.0-1091, 1.0-1092.}

Forrester Research has named Google Cloud a Leader in The Forrester Wave™: AI Infrastructure, Q4 2021 report authored by Mike Gualtieri and Tracy Woo. In the report, Forrester evaluated dimensions of AI architecture, training, inference and management against a set of pre-defined criteria. Forrester’s analysis and recognition gives customers the confidence they need to make important platform choices that will have lasting business impact. 

Google received the highest possible score in 16 Forrester Wave evaluation criteria: architecture design, architecture components, training software, training data, training throughput, training latency, inferencing throughput, inferencing latency, management operations, management external, deployment efficiency, execution roadmap, innovation roadmap, partner ecosystem, commercial model, and number of customers.

We believe that Google’s vision to be a unified data and AI solution provider for the end-to-end data science experience is recognized by Forrester, through high scores in the areas of architecture and innovation. We are focused on building the most robust yet cohesive experience to enable our customers to leverage the best of Google every step of the way. Here are four key areas where Google excels, among the many highlighted in this report. 

AI Infrastructure: Leverage the building blocks of innovation

When an organization chooses to run its business on Google Cloud, it benefits from innovative infrastructure available globally. Google offers users a rich set of building blocks such as Deep Learning VMs and containers, the latest GPUs/TPUs and a marketplace of curated ISV offerings to help architect your own custom software stack on VMs and/or Google Kubernetes Engine (GKE). 

Google provides a range of GPU & TPU accelerators for various use cases, including high performance training, low cost inferencing and large-scale accelerated data processing. Google is the only public cloud provider to offer up to 16 NVIDIA A100 GPUs in a single VM, making it possible to train very large AI models on a single node. Users can start with one NVIDIA A100 GPU and scale to 16 GPUs without configuring multiple VMs for single-node ML training.  Google also provides TPU pods for large-scale AI research with PyTorch, TensorFlow, and JAX.  The new fourth generation TPU pods deliver exaflop-scale peak performance with leading results in recent MLPerf benchmarks which included a 480 billion parameter language model.   

Google Kubernetes Engine provides the most advanced Kubernetes services with unique capabilities like Autopilot, highly automated cluster version upgrades, and cluster backup/restore. GKE is a good choice for a scalable multi-node bespoke platform for training, inference and Kubeflow pipelines, given its support for 15,000 nodes per cluster, auto-provisioning,  auto-scaling and various machine types (e.g. CPU, GPU, TPU and on-demand, spot). ML workloads also benefit from GKE’s support for dynamic scheduling, orchestrated maintenance, high availability, job API, customizability, fault tolerance and ML frameworks.  When a company’s footprint grows to a fleet of GKE clusters, its data teams can leverage Anthos Config Management to enforce consistent configurations and security policy compliance. 

Comprehensive MLOps: Build models faster and more easily without skimping on governance 

Google’s fully managed Vertex AI platform provides services for ML lifecycle management, from data ingestion and preparation all the way up to model deployment, monitoring, and management. Vertex AI requires nearly 80% fewer lines of code to train a model versus competitive platforms1, enabling data scientists and ML engineers across all levels of expertise to implement Machine Learning Operations (MLOps) so they can efficiently build and manage ML projects throughout the entire development lifecycle. 

Vertex AI Workbench provides data scientists with a single environment for the entire data-to-ML workflow, enabling data scientists to build and train models 5x faster than traditional notebooks. This is enabled by integrations across data services (like Dataproc, BigQuery, Dataplex, and Looker), which significantly reduce context switching.  Users are also able to access NVIDIA GPUs, modify hardware on the fly, and set up idle shutdown to optimize infrastructure costs. 

Organizations can then build and deploy models built on any framework (including TensorFlow, PyTorch, Scikit learn or XGBoost) with Vertex AI, with built-in tooling to track a model’s performance.  Vertex Training also provides various approaches for developing large models including Reduction Server to optimize bandwidth and latency of multi-node distributed training on NVIDIA GPUs for synchronous data parallel algorithms.  Vertex AI Prediction is serverless, and performs automatic provisioning and deprovisioning of nodes behind the scenes to provide low latency online predictions. It also provides the capability to split traffic between multiple models behind an endpoint. Models trained in Vertex AI can also be exported to be deployed in private or other public clouds.Google’s strengths in its current offering are in architecture, training, data throughput, and latency. Its sweet spot is in its product offering, Vertex AI, which has core AI compute capabilities and MLOps services for end-to-end AI lifecycle management.The Forrester Wave:™ AI Infrastructure, Q4 2021

In addition to building models, it is important to deploy tools for governance, security, and auditability. These tools are crucial for compliance in regulated industries, and they help teams to protect data, understand why given models fail, and determine how models can be improved. 

For orchestration and auditability, Vertex Pipelines and  Vertex ML Metadata tracks the inputs and outputs of an ML pipeline and the lineage of artifacts. Once models are in production, Vertex AI Model Monitoring supports feature skew and drift detection, alerting data scientists. These capabilities speed up debugging and create the visibility required for regulatory compliance and good data hygiene in general.For explainability, Vertex Explainable AI helps teams understand their model’s outputs for classification and regression tasks. Vertex AI tells how much each feature in the data contributed to the predicted result. Data teams can then use this information to verify that the model is behaving as expected, recognize bias in the model, and get ideas for ways to improve the model and training data.

These services together aim to simplify MLOps for data scientists and ML engineers, so that businesses can accelerate time to value for ML initiatives.

Security: Protect data while keeping ML pipelines flowing  

The Google stack builds security through progressive layers that deliver defense in depth. To accomplish data protection, authentication, authorization and non-repudiation, we have measures such as boot-level signature and chain-of-trust validation. 

Ubiquitous data encryption delivers unified control over data at-rest, in-use, and in-transit, with keys that are held by customers themselves. 

We offer options to run in fully encrypted confidential environments utilizing managed Hadoop or Spark with Confidential Dataproc or Confidential VMs.  

Partner Ecosystem: Work with world-class AI specialists 

Google works with certified partners globally to help our customers design, implement and manage complex AI systems. We have a growing list of partners with Machine Learning specializations on Google who have demonstrated customer success across industries, including deep partnerships with the largest Global System Integrators. The Google Cloud  Marketplace also provides a list of technology partners who allow enterprises to deploy machine learning applications on Google’s AI infrastructure.

Our dedication to being your partner of choice for ML Needs  

Leading organizations like OTOY, Allen Institute for AI and DeepMind (an Alphabet subsidiary) choose Google for ML, and enterprises like Twitter, Wayfair and The Home Depot shared more about their partnership with Google in their recent sessions at Google Next 2021.

Establishing well-tuned and appropriately managed ML systems has historically been challenging, even for highly skilled data scientists with sophisticated systems. With the key pillars of Google’s investments above, organizations can build, deploy, and scale ML models faster, with pre-trained and custom tooling, within a unified AI platform. 

We look forward to continuing to innovate and to helping customers on their digital transformation journey. To download the full report, click here. Get started on Vertex AI, learn what’s upcoming with infrastructure for AI and ML at Google here, and talk with our sales team.

Vertex AI launched with the premise “one AI platform, every ML tool you need.” Let’s talk about how Vertex AI streamlines modeling universally for a broad range of use cases. 

The overall purpose of Vertex AI is to simplify modeling so that enterprises can fast track their innovation, accelerate time to market, and ultimately increase return on ML investments. Vertex AI facilitates this in several ways. Features like Vertex AI Workbench, for example, speed up training and deployment of models by five times compared to traditional notebooks. Vertex AI Workbench’s native integration with BigQuery and Spark means that users without data science expertise can more easily perform machine learning work. Tools integrated into the unified Vertex AI platform, such as state of the art pre-trained APIs and AutoML, make it easier for data scientists to build models in less time. And for modeling work that lends itself best to custom modeling, Vertex AI’s custom model tooling supports advanced ML coding, with nearly 80% fewer lines of code required (compared to competitive platforms) to train a model with custom libraries. Vertex AI delivers all this while maintaining a strong focus on Explainable AI. 

Yet organizations with the largest investments in AI and machine learning, with teams of ML experts, require extremely advanced toolsets to deliver on their most complex problems. Simplified ML modeling isn’t relegated to simple use cases only.

Let’s look at Vertex AI Neural Architecture Search (NAS), for instance. 

Vertex AI NAS enables ML experts at the highest level to perform their most complex tasks with higher accuracy, lower latency, and low power requirements. Vertex AI NAS originates from the deep experience Alphabet has with building advanced AI at scale. In 2017, the Google Brain team recognized we need a better way to scale AI modeling, so they developed Neural Architecture Search technology to create an AI that generates other neural networks, trained to optimize their performance in a specific task the user provides. To the astonishment of many in the field, these AI-optimized models were able to beat a number of state of the art benchmarks, such as ImageNet and SOTA mobilenets, setting a new standard for many of the applications we see in use today, including many Google-internal products. Google Cloud saw the potential of such a technology and shipped in less than a year a productized version of the technique (under the brand AutoML). Vertex AI NAS is the newest and most powerful version of this idea, using the most sophisticated innovation that has emerged since the initial research.

Customer organizations are already implementing Vertex AI NAS for their most advanced workloads. Autonomous vehicle company Nuro is using Vertex AI NAS, and Jack Guo, Head of Autonomy Platform at the company, states, “Nuro’s perception team has accelerated their AI model development with Vertex AI NAS. Vertex AI NAS have enabled us to innovate AI models to achieve good accuracy and optimize memory and latency for the target hardware. Overall, this has increased our team’s productivity for developing and deploying perception AI models.” 

And our partner ecosystem is growing for Vertex AI NAS. Google Cloud and Qualcomm Technologies have collaborated to bring Vertex AI NAS to the Qualcomm Technologies Neural Processing SDK, optimized for Snapdragon 8. This will bring AI to different device types and use cases, such as those involving IoT, mixed reality, automobiles, and mobile.

Google Cloud’s commitments to making machine learning more accessible and useful for data users, from the novice to the expert, and to increasing the efficacy of machine learning for enterprises are at the core of everything we do. With the suite of unified machine learning tools within Vertex AI, organizations can take advantage of every ML tool they need on one AI platform. 

Ready to start ML modeling with Vertex AI? Start building for free. Want to know how Vertex AI Platform can help your enterprise increase return on ML investments? Contact us.