AWS Graviton - AshokBhat/notes GitHub Wiki

Graviton series

  • First-generation - a1
  • Second-generation - m6g, c6g, r6g
  • Third-generation - c7g

Graviton series

Area Graviton A1 Graviton 2 Graviton 3
Instances a1 m6g, c6g, r6g c7g
vCPUs 16 64 64
Process 16nm 7nm 5nm
Transistors ~5 billion ~30 billion ~55 billion
Core [Cortex A72]]](/AshokBhat/notes/wiki/[[Neoverse-N1) Neoverse V1
Memory channels 8 DDR4-3200 channels 8 DDR5-4800 channels
L2 cache per core 0.5MB? 1MB ?
L3 cache 32MB ?
PCIe lanes 64 PCIe4 PCIe5
EBS bandwidth 3.5Gbps 18Gbps
Networking bandwidth 10Gbps 25Gbps
Memory up to 32GB up to 512GB

Performance comparison

Area Graviton A1 Graviton 2
Overall 1x 7x
Per Core 1x 2x
Memory 1x 5x

Graviton2 - EC2 instances

  • m6g - For general purpose
  • c6g - For compute-optimized workloads - HPC, Machine learning inference
  • r6g - For memory-optimized workloads
  • t4g - For burstable general-purpose workloads

Graviton2 - Launch slides

SPEC CPU 2017 ML

Graviton3 - Launch slides

Memory ML Inference vs Graviton2 Bert uplift

See also

  • [AWS SageMaker]] ](/AshokBhat/notes/wiki/[[AWS-Lambda)
  • [Amazon EC2]] ](/AshokBhat/notes/wiki/[[AWS-Elastic-Inference)
  • [AWS Graviton]] ](/AshokBhat/notes/wiki/[[AWS-Inferentia)

Other CPU vendors