AWS re:Invent 2022 Keynote by Peter DeSantis

Here I have captured key highlights from 1+ hour of AWS re:Invent 2022 Keynote by Peter DeSantis, AWS SVP, Utility Computing.

New AWS services announced are preceded by New & bold font

Keynote by Peter DeSantis, AWS SVP, Utility Computing

• AWS Nitro silicon chip:
  • AWS Nitro architecture:
    • Server = customer instance
    • Hypervisor
    • Beneath is Nitro Controller = Networking, Storage, Management security & monitoring
    • Every new EC2 instance introduced since 2014 has Nitro Controller!
  • New AWS Nitro v5 chip launched GA:
    • v5 is Arm based architecture, 2x transistors (means twice the computation), 50% memory bandwidth, 2x PCIe bandwidth
    • New EC2 instance based on v5 = C7gn (Graviton based EC2)
    • New EC2 instance for HPC = HPC7g
• Network topology:
  • Traditional networking based on TCP connection = Single Path through Network
  • SRD = Multi-path through Network (under the hood it uses TCP!), Retries in ms, Runs on Nitro whereas TCP uses Guest operating system!
    • Due to above factors SRD provides Lower Latency & Higher Throughput and is used for EFA (HPC), EBS (storage), ENA (networking)
    • Coming soon - All new EBS io2 volumes will be running on SRD
    • Elastic Network Adapter (ENA) - Default network interface for Amazon EC2
      • ENA Express - bringing SRD to mainstream
• ML: Training performance of Neural Network Models:
  • Stochastic Rounding resulted in fast convergence as compared to 32 bit or 16 bit or Mixed precision
  • Upto 75% Faster synchronization with Ring of Rings algo
  • New Trn1n (Trainium) based EC2 instances - coming soon
• Serverless computing:
  • AWS Lambda: launched in 2014 - first serverless computing service
    • From t2 EC2 instance per Lambda function to FireCracker (micro-VM) to avoid Cold start (reduced by 50%)
    • Cold start occurs due to Compute Cache Hits
    • New Lambda SnapStart for Java: significantly reduces Lambda function cold start latency, at no extra cost - in GA
      • Upto 90% reduction in cold start latency
      • Initialization snaphot taking for JVM & used across different Lambda functions!
      • No code change required, just enable for new or existing Java Lambda
• Ferrari Racing session by Jock Clear

References

• Watch now on YouTube: Keynote by Peter DeSantis

• Explore: AWS re:Invent 2022