AWS EC2 R9g and R9gd Are Now GA: Graviton5's Case for Memory-Optimized Workloads

Graviton5 has been available in compute-optimized instances for months. With the R9g and R9gd families going GA, it reaches the workloads where memory bandwidth is the bottleneck — and where the specification changes matter most.

What Happened

Amazon EC2 R9g and R9gd instances are now generally available, powered by AWS Graviton5 processors. Eleven sizes are available, from medium (1 vCPU, 8 GiB) to metal-48xl (192 vCPUs, 1,536 GiB). Network bandwidth tops out at 100 Gbps on the 48xlarge; EBS bandwidth reaches 72 Gbps.

R9gd variants add local NVMe SSD storage (59 GB on smaller instances to 3,800 GB on the largest) for workloads that need fast local scratch or high-frequency write throughput alongside large memory.

Current availability spans US East (N. Virginia, Ohio), US West (Oregon), and Europe (Frankfurt), via On-Demand, Spot, Savings Plans, and Dedicated options.

Why It Matters

Three Graviton5 hardware improvements are specifically relevant for memory-optimized use cases:

DDR5 at 8,800 MT/s. AWS describes this as “the fastest memory available in the cloud.” For workloads where cache hit rates and memory bandwidth determine throughput — in-memory databases, hot-tier Redis or Valkey deployments, real-time analytics — the bandwidth ceiling directly affects query latency and concurrency.

5x larger L3 cache than Graviton4. Cache expansion compounds the DDR5 improvement: more working set fits in L3, reducing the frequency of DRAM accesses. For databases with predictable hot-data access patterns, this is a meaningful architecture change rather than an incremental spec bump.

Up to 25% better compute performance versus R8g. Memory-optimized instances are often CPU-bound on the query execution path even when memory bandwidth is the nominal constraint. The compute improvement means the transition from R8g doesn’t require a careful tradeoff between compute and memory — both improve.

Network and EBS bandwidth are also up to 2x higher and packet processing up to 3x higher than Graviton4, which matters for database replication, cross-zone traffic, and EBS-backed write-intensive workloads.

What This Means for Builders

The primary use cases AWS targets: managed databases, in-memory caching (Valkey, Redis, Memcached), real-time analytics, containerized services, and Kubernetes deployments across multiple runtimes.

For teams running RDS or ElastiCache on R8g, the GA announcement is the trigger to begin right-sizing analysis. Graviton-optimized instances have historically offered better price-performance for memory-intensive workloads compared to Intel and AMD alternatives — the R9g extends that advantage.

For Kubernetes deployments with memory-heavy pods (JVM services, large language model inference with significant attention cache requirements), the R9gd variant’s local NVMe is worth evaluating for ephemeral scratch volumes that would otherwise pull IOPS from EBS.

The 5x L3 cache expansion deserves specific attention for database engine operators: the effective hot-data cache may shift workloads from memory-bound to compute-bound, changing the optimal instance size. Benchmarking your specific workload against R8g is more informative than extrapolating from the 25% headline figure.

Source: AWS Blog

Content created with AI assistance and reviewed for accuracy.