Google Cloud has moved the Z4D storage‑optimized machine family to general availability, adding both virtual‑machine and bare‑metal offerings that deliver up to 84 TB of local SSD, up to 384 vCPUs, and 5th‑generation AMD EPYC (Turin) processors. The higher SSD density, up to 70 % I/O performance gains and doubled network bandwidth give AI, data‑analytics and database engineers a way to increase throughput without over‑provisioning compute.
Hardware specs and raw performance
The Z4D instances are built on Titanium SSDs, which shift local storage processing off the CPU. Reported capabilities include:
- Maximum of 84,000 GiB local SSD per instance.
- Up to 15,600K random‑read IOPS and 75,600 MiB/s sequential‑read throughput.
- Performance improvements of up to 70 % over the previous Z3 generation.
- Write latency reduced by as much as 25 % and mixed read‑write IOPS up by 30 % without increasing overall I/O latency.
- Standard networking bandwidth up to 400 Gbps, roughly double the Z3 offering.
VM types and workload fit
Two VM families address different storage‑intensity patterns:
- Z4D‑highmem‑standardlssd: seven shapes, 219 GiB local SSD per vCPU, tuned for OLAP workloads and relational databases such as MySQL and PostgreSQL.
- Z4D‑highmem‑highlssd: seven shapes, 438 GiB local SSD per vCPU, aimed at distributed databases, high‑throughput data streaming, large parallel file systems, and search workloads.
Both families can be mixed into existing Z3 clusters, allowing incremental migration and capacity scaling.
Bare‑metal considerations
Bare‑metal Z4D instances expose the physical hardware directly, eliminating the hypervisor layer. This reduces latency for latency‑sensitive workloads, supports custom hypervisors, and satisfies licensing models that require bare‑metal access. They are the first AMD‑based instances to support Nutanix Cloud Clusters (NC2), a hybrid multi‑cloud platform. The low‑latency, high‑capacity local SSD makes them suitable for agentic AI architectures that run thousands of microVM sandboxes per host.
Operational and security implications
Offloading storage processing to the Titanium SSDs can free CPU cycles for application workloads, but monitoring stacks should be updated to capture SSD‑level metrics rather than CPU‑bound storage metrics. The advertised “enhanced storage security” suggests that data at rest benefits from the SSD design, though specific controls are not detailed. The increase in network bandwidth may require revisiting egress budgeting and firewall rules. Bare‑metal licensing and hypervisor choices should be validated against organizational compliance policies.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
- Profile your workloads for I/O intensity; if local SSD throughput or capacity is a bottleneck, evaluate Z4D VM shapes that match your per‑vCPU SSD ratio.
- Run targeted benchmarks (e.g., indexing, random read) to confirm the claimed 40‑50 % throughput gains for your specific database or AI training pipeline.
- Plan for the higher network ceiling; ensure that downstream services and load balancers can handle up to 400 Gbps without saturating.
- Consider bare‑metal Z4D if you need sub‑microsecond latency, custom hypervisors, or Nutanix Cloud Cluster integration.
- Update observability dashboards to include Titanium SSD metrics and verify that security policies cover the new storage layer.