Amazon has officially launched its new EC2 G7 instances, marking the first time a major cloud provider has deployed NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs for public cloud workloads. These instances, designed for AI inference, graphics rendering, and data analytics, provide up to 4.6 times the AI inference performance compared to the previous G6 generation, according to official AWS documentation.
Technical Specifications and Performance Gains
The G7 instances represent a significant architectural shift from the preceding G6 generation. Each instance is powered by the NVIDIA Blackwell architecture, which features 5th Gen Tensor Cores and 4th Gen RT Cores. According to AWS, these chips provide 1.33 times the memory capacity and 2.45 times the memory bandwidth of G6 hardware.
Networking throughput has also seen a substantial upgrade, reaching 700 Gbps via EFA-enabled connectivity. This represents a seven-fold increase over G6 networking capabilities, intended to reduce latency in multi-node clusters. The integration of ninth-generation NVENC and sixth-generation NVDEC engines allows for a 1.5x increase in concurrent high-resolution video streams, supporting complex 4:2:2 encoding workflows.
Comparison: G7 vs. G6 Instances
The following table highlights the key performance improvements reported by AWS regarding the transition to G7 hardware:
| Feature | G6 Instance | G7 Instance |
|---|---|---|
| AI Inference Performance | Baseline | Up to 4.6x faster |
| Graphics Performance | Baseline | Up to 2.1x faster |
| Network Throughput | 100 Gbps | Up to 700 Gbps |
| Video Stream Capacity | Baseline | 1.5x throughput |
Deployment and Availability
AWS has made G7 instances available immediately in two regions: US East (Ohio) and US West (Oregon). Users can deploy these instances using AWS Deep Learning AMIs or NVIDIA Workstation AMIs. For those utilizing Amazon EKS, the company requires the use of NVIDIA driver version R595 to ensure compatibility with the Blackwell architecture.
The instances support standard operating systems, including Amazon Linux, Ubuntu, RHEL, and Windows Server. By incorporating support for industry-standard graphics libraries—such as DirectX, Vulkan, and OpenGL—AWS aims to capture enterprise demand for virtual desktop infrastructure (VDI) and spatial computing.
Pricing and Purchasing Options
Organizations can access G7 instances through several billing models:
- On-Demand: Standard hourly billing for compute time.
- Savings Plans: Long-term commitments for reduced rates.
- Spot Instances: Excess capacity offered at a lower price point for fault-tolerant workloads.
Dedicated instance hosting is currently available for the 12xlarge, 24xlarge, and 48xlarge configurations. Pricing details for these specific tiers are available via the official Amazon EC2 Pricing page.
Why This Matters for AI Infrastructure
The shift to Blackwell-based instances reflects a broader industry trend toward hardware specialized for inference rather than just training. While training models requires massive, interconnected GPU clusters, inference—the process of running a model to generate predictions—demands high memory bandwidth and low-latency throughput to maintain user-facing response times. By providing 32 GB of GPU memory per unit and up to 256 GB total in the largest configurations, AWS is positioning the G7 series as a primary choice for enterprises scaling generative AI applications into production environments.