Google Cloud Introduces Z4M Storage-Optimized Compute Instances for AI/ML Workloads

New Z4M Machine Series for AI/ML

Google Compute Engine has announced the general availability of its Z4M machine series. These new storage-optimized compute instances are available as both Virtual Machine (VM) and bare-metal instances. Z4M is designed for AI/ML training and inference workloads that require high storage and networking throughput, such as checkpointing, distributed key-value caches, and parallel file systems.

Hardware Specifications and Performance

Z4M instances are powered by fifth-generation Intel Xeon Scalable processors and incorporate Titanium enhancements. They offer up to 168,000 GiB of Local SSD (LSSD) storage and up to 400 Gbps of Cloud Remote Direct Memory Access (RDMA) networking. RDMA reduces data-access latency by up to 60% compared to TCP networking, which helps maintain GPU and TPU utilization.

Z4M VMs and bare-metal instances provide up to 192 vCPUs and 1,536 GiB of memory. The Titanium SSDs offload local storage processing from CPU resources. Z4M delivers up to 15.6M random-read IOPS and up to 75,600 MiB/s sequential-read throughput, representing up to a 70% improvement in LSSD storage performance and 110% in throughput compared to the Z3 series.

VM and Bare-Metal Options

The Z4M VM portfolio allows scaling infrastructure with large total LSSD capacity and 875 GiB of LSSD capacity per vCPU. Z4M bare-metal instances provide direct access to physical hardware, bypassing the virtualization layer. This reduces latency for workloads like custom hypervisors and those with specific licensing requirements. Bare-metal instances are also suitable for agentic AI workloads, enabling isolated sandboxes at scale.

Networking and Data Access

Z4M utilizes 400 Gbps of RDMA-enabled standard networking to provide low-latency access to the storage layer. This feature helps ensure GPUs and TPUs remain fully utilized, even when processing large multimodal datasets, by minimizing data access bottlenecks.