This is one fixed Valkey 9.0 load read across six finished Amazon ElastiCache reports: cache.t3.micro and cache.t4g.micro, cache.t3.small and cache.t4g.small, cache.t3.medium and cache.t4g.medium. Two axes at once: Intel (T3) versus Graviton (T4g), and micro to small to medium scaling.

The result is not a flat winner. Micro and small nodes reached maxmemory, so they read as behavior under memory pressure. Only the medium pair kept real headroom. Underneath the headline ops/s sit three signals most comparisons skip: burstable CPU credits, engine-side latency versus client-observed latency, and how much each node swapped.

The Valkey 9.0 ElastiCache cache.t3.small benchmark on cache.t3.small now has a real load-generator-side latency signal. The earlier latency view was an empty placeholder, this run's ECS artifact includes Embedded Metric Format samples for ElastiCache/LoadGenerator / ClientLatency with p50, p99, and p99.9.

The cache.t4g.micro result set currently splits into two useful signals: Valkey 7.2 leads throughput and client-side average latency, while Valkey 9.0 shows the lowest throughput variation and peak engine CPU. Redis 7.1 stays in the set as the Redis reference run.