DEV Community

Manu Muraleedharan
Manu Muraleedharan

Posted on

Cool Announcements @ AWS ReInvent CEO KeyNote

Image description

S3 Express One Zone
Highest performance lowest latency cloud storage
50% less cost than S3 standard
10 times faster than S3 standard
Single-digit millisecond latency
Millions of requests per second
Co-locate storage and compute in the same AZ
Pinterest has 10x faster write-speed and 40% less cost from S3 Express One Zone

Graviton 4
30% faster than Graviton3
Faster and more energy-efficient
R8g Ec2 instance with Graviton4 in preview

EC2 Ultra-clusters
20,000 GPUs, connected by EFA of 32000 GBPS
Equal to a supercomputer = 20 Exaflops

New GPU = GH200 which is 4 times faster with new LLM compilers
Uses Grace Hopper technology to connect CPU and GPU at 1TBPS
32 GH200 can connect via the NVLINK switch.

NVIDIA DGX cloud is coming to AWS
This is NVIDIA's AI Factory, connecting 16,000+ GPUs together.
65 Exaflops compute capacity and will make LLM learning 2x fast

*EC2 Capacity Blocks for ML *
Reserve EC2 Ultra Clusters for short-term usage

Trainium 2
4x faster, second-gen specially designed chips for training models

ML Customization in AWS

Fine-tuning available in:
Titan Text Lite, Express, Cohere Command Lite, Meta Llama 2, and Anthropic Claude

Retrieval-Augmented Generation with Knowledge Bases
(Announced Sept 2023)

Continued Pre-Training is available in AWS Bedrock. The technique involves using large amounts of unlabeled data before fine-tuning a model,

*Agents for Bedrock *
Execute multi-step actions across company systems powered by ML

Guardrails for Bedrock
Safeguard generative AI applications with responsible AI policies

Education Commitment
AWS commits to training 29 million people in the cloud and 2 million in AI for free by 2025.

*AWS CodeWhisperer Customization Capability *
Provide custom code suggestions using internal SDK, api, and code.

Amazon Q (Some of the features in Preview)
This is the new service I am most excited for.
AI assistant designed for the business world that understands your company information.
Can chat with Q in AWS console, documentation, code whisperer, and chat apps like Slack.
Q is already trained with all the AWS information, WAF principles, etc.
Troubleshoot errors with Q when you get an error in the AWS console.
Get recommendations and step-by-step information
Feature Development: Develop a new feature in AWS using Q using prompts interactively and iteratively
Code Transformation: Use Q to upgrade language versions in code eg: 1000 Java apps upgraded in 2 days
Business expert: Connect to over 40 data sources and answer business questions, supports RBAC
Amazon Q inside Quicksight: Create BI reports and visualizations using generative AI by Q
Amazon Q inside Amazon Connect: AI Agent stays on-call, to help on-call agents with customer interaction

Zero-ETL integrations with Redshift in:
Aurora Postgres, RDS for MySQL, DynamoDB
As soon as data is written into these databases, query and analyze in Redshift without ETL pipeline.

Zero-ETL integration between DynamoDB and OpenSearch
Search DyanmoDB data through Opensearch without doing ETL.

AI recommendations for Amazon DataZone
Add business descriptions to data in DataZone using AI.

Project Kuiper
A constellation of low Earth orbit satellites that aims to provide fast, affordable, and reliable broadband to customers in areas without reliable internet connection. Private network connectivity is now available.

Top comments (0)