Graviton5's 25% per-vCPU gain builds the cost case for pricier R9g reservations

The gain has a shape. Graviton5 pairs the faster DDR5 with an L3 cache five times the size of Graviton4's ^[4], and the memory transfer rate alone works out about 57% higher ^[19]. That points at a specific kind of code: working sets that spilled out of the old L3 and now fit, and scans that stream from DRAM and were waiting on it. A service whose tail latency is set by a downstream call or a contended lock keeps that tail regardless of how fast the local memory is.

AWS also publishes per-segment figures: up to 30% faster for databases, up to 35% faster for web applications and up to 35% faster for machine learning ^[9]. Ranking web applications above databases on a memory-optimised part is worth a raised eyebrow, and no methodology accompanies either number ^[9]. For the database figure to transfer, your database would need to be bound on compute and memory in roughly the same proportion as the one AWS measured, which for most fleets means first checking whether it is actually waiting on EBS.

Then the arithmetic that decides a commitment. If a vCPU does 25% more work, the same throughput needs 1/1.25 of the vCPUs, which is 20% fewer ^[20]. That is the break-even: R9g can carry up to a 25% higher price per vCPU-hour and still land at the same cost per unit of work ^[20]. Neither AWS announcement states an hourly price for R9g or R9gd ^[22], so the "best price performance for memory-intensive workloads" line ^[17] cannot be checked against the launch material that makes it.

The adoption cost is the part that is documented. Multi-arch container images built for Arm64 run without changes ^[13], and R9gd matches R9g on compute and networking with local NVMe added ^[23], so an R8g operator is buying a resize and a load test rather than a port. One thing here is plain good engineering: the Nitro Isolation Engine mediates virtual machine memory, register state and I/O through a minimal API set, and AWS says formal verification of that component makes Nitro the first formally verified cloud hypervisor ^[14].