Skip to main content

AWS Inferentia 2 (Inf2)

Overview

AWS Inferentia 2, released in 2022, is the second-generation AWS custom inference ASIC. The major improvement is support for 32GB HBM2e memory and inter-chip interconnect, enabling hosting of larger models (70B+). It is available through Inf2 instances, ranging from inf2.xlarge (single chip) to inf2.48xlarge (12 chips).

Core Specifications

ItemSpec
ArchitectureInferentia 2 (NeuronCore-v2)
NeuronCore4 per chip
Memory32 GB HBM2e
INT8 Compute (estimated)190 TOPS
FP16/BF16190 TFLOPS (estimated)
Inter-Chip InterconnectSupported (up to 12 chips)
TDP~150 W

Inf2 Instances

InstanceChipsMemoryUse
inf2.xlarge132 GBLightweight inference
inf2.8xlarge132 GBModerate inference
inf2.24xlarge6192 GBLarge models
inf2.48xlarge12384 GB70B+ models

Vendor Information

ItemDetail
ManufacturerAmazon (Annapurna Labs)
Websitehttps://aws.amazon.com/machine-learning/inferentia/
Product Pagehttps://aws.amazon.com/ec2/instance-types/inf2/
SDKhttps://awsdocs-neuron.readthedocs-hosted.com/

Key Features

  • Inter-chip interconnect: Supports 12 chips aggregating 384GB memory
  • Can host 70B model inference (4-bit quantization)
  • DynamoAI integration
  • Low latency

Use Cases

  • LLM inference (10B-70B)
  • Recommendation systems
  • Computer vision inference
  • Cost-sensitive production deployment