Introducing Z4M: storage-optimized compute for AI/ML and parallel file systems
Google Cloud launches Z4M machine series for AI/ML and parallel file systems, offering up to 168,000 GiB of Local SSD and 400 Gbps RDMA networking to reduce data-access latency by up to 60%.
Google Cloud has introduced the Z4M machine series in Compute Engine, available as both virtual machines and bare-metal instances, designed for AI/ML workloads requiring high storage and networking throughput. The series leverages fifth-generation Intel Xeon Scalable processors and Titanium SSDs to deliver up to 168,000 GiB of Local SSD storage and 400 Gbps Cloud RDMA networking, reducing data-access latency by up to 60% compared to TCP. This setup aims to keep GPUs and TPUs fully utilized during large-scale data processing tasks.
Z4M instances provide up to 192 vCPUs and 1,536 GiB of memory, with storage capacities ranging from 875 GiB per vCPU in VMs to 168,000 GiB in bare-metal configurations. The series supports GPUDirect Storage and TPUDirect Storage, enabling line-rate data transfers to maintain high throughput for multimodal datasets. Titanium SSDs offload storage processing from CPUs, delivering up to 15.6M random-read IOPS and 75,600 MiB/s sequential-read throughput, representing a 70% and 110% improvement over the previous Z3 series, respectively.
The Z4M VM portfolio allows scaling infrastructure to meet workload demands, with options for large Local SSD capacity and balanced LSSD-to-networking throughput ratios. Bare-metal instances provide direct hardware access, reducing latency for custom hypervisors and agentic AI workloads, including support for Agent Substrate. Both VMs and bare-metal instances offer advanced maintenance scheduling, including live migration for smaller storage configurations and planned termination for larger setups while preserving data.
Z4M instances are compatible with Hyperdisk storage options, including Balanced, Throughput, and Extreme tiers, supporting up to 512 TiB of capacity per instance. Hyperdisk Extreme delivers up to 500K IOPS and 12,500 MiB/s throughput, while Balanced offers up to 320K IOPS for cost-efficient general-purpose workloads. The series is available in select regions, with deployment options via the Google Cloud console or GKE node pools.