No benchmarks yet, but I can get 30% faster latency, on 2/3 the number of instances compared to c6g for the same uncompress/compress/write to Kafka workload.