Google Cloud Z4M Pairs Dense Local SSD Storage With 400 Gbps RDMA

Nov 06, 2026 479 views

For data-heavy AI systems, storage capacity alone doesn’t keep accelerators busy. The harder problem is feeding data to them quickly while handling checkpoints, caches and large datasets at scale. Google Cloud’s Z4M storage-optimized machine series targets that bottleneck with up to 168,000 GiB of local SSD and 400 Gbps of RDMA networking, in both virtual machine and bare-metal configurations.

The series is now generally available in select regions worldwide. Built on fifth-generation Intel Xeon Scalable processors and Google’s Titanium infrastructure, Z4M combines high local storage density with direct, low-latency data movement. Google says RDMA can reduce data-access latency by up to 60% compared with TCP networking. For AI training and inference, the practical aim is to keep GPUs and TPUs supplied with data rather than waiting on storage.

Storage density meets higher throughput

Z4M instances offer up to 192 vCPUs and 1,536 GiB of memory, with as much as 875 GiB of local SSD per vCPU. Google reports peak local storage performance of 15.6 million random-read IOPS and 75,600 MiB/s of sequential-read throughput. Compared with the previous Z3 generation, those figures represent increases of up to 70% in storage performance and 110% in throughput, according to the company.

The local drives are Titanium SSDs. Google says Titanium shifts local storage processing away from CPU resources, supporting low-latency, high-throughput access and enhanced storage security. That offload matters when a machine is doing more than serving a single application: storage-intensive work can compete with compute for CPU time, while AI pipelines and databases depend on steady data delivery.

Z4M’s 400 Gbps RDMA-enabled networking is designed to move data between storage and accelerators at line rate. The platform supports NVIDIA GPUDirect Storage and TPUDirect Storage, which enable direct paths between storage and GPU or TPU memory. Google describes the balance between local SSD and network throughput as a 1.4:1 ratio, intended to let instances process data locally while sending it to clients at full network speed.

That balance broadens Z4M’s use beyond model training. The company positions the machines for checkpointing and distributed key-value caches, as well as parallel file systems and enterprise storage. Its stated use cases also include SQL and NoSQL databases, vector databases and search workloads, where sustained I/O can support high transaction volumes.

VMs and bare metal serve different needs

The VM configurations provide a way to scale infrastructure and clusters around workload requirements while retaining high local SSD capacity. Bare-metal instances expose physical hardware without a virtualization layer, a design Google says can reduce latency for custom hypervisors and workloads with specific licensing requirements. The company also points to isolated sandbox environments for agentic AI development, including use with Agent Substrate.

Storage vendors cited performance results from their own testing. Silk reported up to 3.7 times lower latency and up to six times lower CPU usage with Cloud RDMA on Z4M compared with TCP. WEKA said its tests found roughly twice the read throughput and up to 79% more IOPS over the Z3 family. These are vendor-reported results, and workloads will determine how closely individual deployments match them.

The launch also includes support from Nutanix, Sycomp, Silk and WEKA across virtualization, data storage and parallel file system use cases. Nutanix says it plans to bring Nutanix Cloud Clusters (NC2) to Z4M bare-metal instances. Its vice president of product management, Saveen Pakala, described the combination as suited to dense virtual desktop infrastructure clusters and data-intensive AI workloads, citing Z4M’s I/O performance and storage density alongside Intel compute and AMX efficiency.

Sycomp plans to support Z4M through its Intelligent Data Storage Platform. Senior HPC Solutions Architect Scott Fadden pointed to greater capacity per node and the continuing development of the Z-series architecture. The partner statements indicate that Z4M is intended to fit into existing storage and infrastructure stacks, as well as serve as a Google Cloud machine type.

Cluster placement and maintenance choices

Google’s Cluster Director and Cluster Toolkit are available for configuring and deploying Z4M clusters. Cluster Director supports placement and maintenance planning: operators can choose a compact topology to prioritize performance or distribute instances across availability domains for higher availability. Maintenance can be coordinated or staggered across those domains. Z4M instances can also be placed alongside GPU and TPU clusters to process large datasets near the accelerators.

Maintenance behavior depends on the instance configuration. Google says it will provide several days’ notice before required maintenance. Z4M VMs with up to 42,000 GiB of local SSD can live-migrate during maintenance. VMs with at least 84,000 GiB of local SSD, along with bare-metal instances, must be stopped and restarted; Google says data is preserved through planned maintenance events.

Network-attached storage remains configurable

Z4M also supports Hyperdisk, Google Cloud’s block storage service, so customers can tune attached storage capacity and performance independently of the instance’s local SSD. Supported options include Hyperdisk Balanced, Hyperdisk Throughput and Extreme Hyperdisk. Together, they support up to 512 TiB of network-attached capacity per instance.

For general-purpose workloads, Hyperdisk Balanced offers up to 320,000 IOPS per instance. Hyperdisk Extreme is aimed at demanding database workloads, with up to 500,000 IOPS and 12,500 MiB/s of throughput per Z4M instance. Those options let operators pair local storage for fast, nearby data access with network-attached capacity sized to a workload’s broader needs.

Availability and deployment

Z4M VMs and bare-metal instances are available in select regions. Customers can choose the Z4M family when creating a VM or a Google Kubernetes Engine node pool in the Cloud console. Regional availability and machine details are listed on the Z4M machine series page; Google also directs customers to its sales team for regional information.

The clearest fit is for infrastructure where local data capacity and network movement are both limiting factors: AI pipelines, parallel file systems and I/O-heavy databases. Z4M’s headline figures establish the available scale, while the choice between VMs and bare metal, local SSD and Hyperdisk, and compact or distributed cluster placement gives teams room to shape deployments around latency, capacity and maintenance requirements.

Source: bob Napaa · cloud.google.com

Comments

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

Introducing Z4M: storage-optimized compute for AI/ML and ...