While the shift of paradigm in machine learning (ML) has been brewing for decades, today’s diverse customer base across industries is reinventing their business models through the maturing of ML technologies, the availability of affordable and abundant scalable computation, and the ubiquity of data.
Generative AI apps such as ChatGPT have gained a lot of interest and attention recently. We think that most consumer experiences and applications will be completely redesigned with generative AI, and we are really at an exciting turning point in the general usage of ML.
For more than 20 years, Amazon has prioritized AI and ML, and ML is the driving force behind many of the features that customers utilize on the platform. Machine learning powers our e-commerce recommendation engine, our supply chain, forecasting, and capacity planning, as well as the paths that optimize robotic picking routes in our fulfillment centers.
Deep learning is employed by Prime Air (our flying drones) & Amazon Go (a grocery store where customers pick items off a shelf, and walk out of the store without the need to pay at a cash register or checkout line). Alexa uses more than thirty different machine learning approaches; it helps half a billion users weekly with shopping, household management, information search, entertainment, and much more.
At Amazon, machine learning is a major component of our history, present culture, and future, with hundreds of engineers dedicated to the field.
At AWS, we have been instrumental in democratizing machine learning (ML) and enabling over 100,000 clients across various industries and sizes to leverage it. At all three tiers of the stack, AWS offers the most extensive and varied range of AI and ML services.
Amazon investments and made innovations to provide the most efficient, scalable infrastructure for ML training and inference at a reasonable price. We've also created Amazon SageMaker, the simplest tool for developers to create, train, and implement models, and Amazon introduced a number of services that let users integrate artificial intelligence features like image recognition, forecasting, and intelligent search into their applications with just an API call.
For this reason, ML is being used by clients like Intuit, Thomson Reuters, AstraZeneca, Ferrari, Bundesliga, 3M, and BMW, in addition to hundreds of startups and government organizations worldwide, to revolutionize their businesses and missions.
Amazon approaches generative AI workloads with the same democratizing mindset, trying to get these technologies out of the lab and into the hands of more people than just a few well-funded startups and major tech giants.
Whether clients are attempting to run, construct, or customize FMs, they require the most efficient, reasonably priced infrastructure specifically designed for machine learning. With our AWS Trainium and AWS Inferentia chips, which offer the lowest cost for training models and conducting inference in the cloud.
AWS has been striving towards the boundaries of performance and affordability for demanding generative AI workloads with AWS EC2 and S3 like ML training and inference over the last five years. Leading AI businesses such as AI21 Labs, Anthropic, Cohere, Grammarly, Hugging Face, Runway, and Stability AI run on AWS because of this platform's ability to optimize performance and minimize costs by selecting the best ML infrastructure.
While 800 Gbps may seem like a lot of bandwidth, we haven't stopped innovating to give you even more. Today, we're excited to announce the general availability of new Trn1n instances that have been network-optimized. These instances offer 1600 Gbps of network bandwidth and are made to provide 20% better performance than Trn1 for large, network-intensive models.
The ease of customizing a model is among Bedrock's most significant features. To fine-tune the model for a specific activity, customers only need to direct Bedrock at a small number of labeled examples in Amazon S3. This eliminates the need to annotate enormous volumes of data—as little as 20 instances would suffice.
Consider a content marketing manager at a top fashion shop who is tasked with creating campaign material and new, targeted advertisements for an impending handbag line. In order to accomplish this, they provide Bedrock a few tagged samples of their most successful taglines from previous campaigns, along with the corresponding product descriptions.
Bedrock educates this specific copy of the model by making a particular copy of the fundamental base model that is exclusive to the customer. After training, Bedrock shall embark on generating efficient messages in the web text, displays ads and social media posts for the new handbags. Customers can be sure that their data will stay private and confidential because none of it is used to train the original base models and is encrypted and stays inside the Virtual Private Cloud (VPC) of the customer.
The majority of the time and money spent on FMs these days is on their training. This is a result of the fact that many clients are only now beginning to employ FMs in production. But when FMs are implemented widely in the future, the majority of the expenses will come from executing the models and conducting inference.
A production application can continuously generate predictions, or inferences, possibly producing millions of predictions per hour, whereas a model is usually trained on a periodic basis. Furthermore, real-time prediction necessitates very low latency and high throughput networking. One excellent example is Alexa, which receives millions of queries per minute and is responsible for 40% of compute costs.
Now, what Amazon is introducing is the availability of Inf2 instances, fueled by the AWS Inferentia2, for public use and these new Inf2 instances are built for gigantic generative AI tasks which in turn involves hundreds of billions of parameters. As compared to the previous first generation Instances built on Inferentia, the second-generation Inf2 instances come with higher –up to 4x – throughput and orders – up to 10 times- lower latency
In addition, they provide extremely fast connection amongst accelerators to facilitate distributed inference on a huge scale. These features result in the lowest cost for inference in the cloud and up to 40% better inference pricing performance compared to other similar AWS EC2 instances. For some of their models, customers such as Runway are experiencing up to two times greater throughput with Inf2 compared to equivalent Amazon EC2 instances.
Runway will be able to add more features, implement more sophisticated models, and ultimately provide a better experience for the millions of creators that use Runway because to this high-performance, low-cost inference.
Thus, machine learning and generative AI workloads are now reshaping industries around the globe. Luminaries such as Amazon and AWS are at the fore, innovating and introducing humongous Artificial intelligence tools leveraging a balanced economy with the help of commoditization and distinctive services such as AWS Trainium, Inferentia, and SageMaker.
With its constant developments, these technologies may contribute to altering consumers’ experience and upgrading organizational processes in the future. It is probably one of the best transformations of ML across the different fields of business.
1.What is AWS Trainium?
AWS Trainium is a chip designed for cost-effective and efficient ML model training in the cloud, offering high performance at a lower cost compared to traditional alternatives.
2. What are AWS Inferentia and Inf2 instances?
AWS Inferentia is a chip optimized for high-performance ML inference, while Inf2 instances are powered by AWS Inferentia2, offering significantly improved throughput and reduced latency for large-scale generative AI applications.
3. What is Amazon SageMaker?
Amazon SageMaker is a fully managed service that enables developers and data scientists to quickly build, train, and deploy machine learning models at scale.
4. How is machine learning used at Amazon?
Machine learning powers various Amazon services including e-commerce recommendation engines, supply chain optimization, robotic picking routes in fulfillment centers, Prime Air (drone delivery), Amazon Go (cashier-less stores), and Alexa (voice assistant).
5. How is AWS democratizing machine learning?
AWS offers a comprehensive suite of AI and ML services accessible to over 100,000 clients across different industries, providing scalable infrastructure, tools like SageMaker for model development, and advanced AI capabilities through simple API integration.