Fastest AI Data Platform in the Cloud

Same cloud, same performance. Half the cost.

Why Legacy File Systems Strain Cloud AI

AI workloads are increasingly being run both on-premises and in the cloud. As compute demands surge— power, cooling, and hardware availability are constrained—driving organizations to increasingly rely on hybrid and cloud computing.

Legacy file systems like Lustre were used in managed cloud services for AI mainly because better options didn’t exist at the time. While they can deliver strong performance for certain workloads, they require significant tuning, specialized expertise, and ongoing management. These systems were built for tightly controlled, specific hardware and networking setups, making them a poor fit for the flexible and diverse environments of the cloud. As a result, their infrastructure needs often conflict with cloud deployment models, leading to wasted resources and higher costs.

Hammerspace was made for this moment – fast GPU I/O speed, easy to manage and scale, cloud-native, and standards-based.

Hammerspace Tier 0 on the Cloud

Hammerspace Tier 0 is a new tier of ultra-fast, shared storage that uses the local NVMe storage in GPU servers and turns it into a tier of shared storage by making those storage volumes part of the Hammerspace Parallel Global File System. Tier 0 can be orchestrated to move seamlessly between tiers of storage, or between sites and clouds.

Hammerspace Tier 0 leverages our standards-based parallel file system architecture – and takes advantage of an update to the Linux kernel that bypasses the NFS client and server, along with the networking stack and network adapter hardware that connects the NFS client and server (depicted in Figure 2 below). Effectively this creates a ‘shortcut’ or more direct data path between the GPUs and the local NVMe storage, which reduces latency, increases bandwidth and as demonstrated in this analysis speeds up checkpointing up by orders of magnitude.

Even with expensive 400Gb/s InfiniBand and high-performance external storage, writing across the network is 2.5x slower than writing directly to local NVMe inside the GPU server. On 100Gb/s networks, it’s nearly 10× slower. Tier 0 provides infrastructure efficiency with the most capacity per unit of bandwidth.

Checkpointing in the Cloud – up to 100X+ Faster

Leveraging Tier 0 In the cloud illustrates an extreme gap in performance. A 500GB checkpoint on an eight-GPU H100 instance takes:

~4.5 seconds to local NVMe (Tier 0)
~139 seconds to AWS EBS io2
~9 minutes to AWS EBS gp3

Async checkpointing helps reduce GPU idle time, but it doesn’t fix the core problem: all data traffic suffers when it has to cross the network. And as context lengths grow and clusters scale, this traffic grows exponentially.

Yes, you can try to match Hammerspace Tier 0 performance with exotic gear—like 800Gb/s networking and expensive storage instances —but at a massive cost in dollars and watts.

Hammerspace gives you that performance for free, using the NVMe already available in the GPU instances you are using.

Training and Inference with Tier 0 - 2.5x Faster

Tier 0 Provides ~2.5X More Read Bandwidth, ~2X More Write Bandwidt

For AI training and GenAI, speed isn’t optional—it’s the bottleneck between models that ship and models that stall. Hammerspace Tier 0 gives you 2.5× faster performance in the cloud than Tier 1 storage because it lives where the GPUs live—inside the server.

Analysis based on benchmark testing using Flexible 1/0 (fio) tester, Clients: 2, Files: 16. Filesize: 50GB. Direct: True. Block size: IMB. IO Depth: 2, IO Engine: libaio. Number of Jobs: 1 (Per File). Run Time=300, Workloads: 100% Sequential Read, 100% Sequential Write, 100% Sequential 50/50 Read/Write Mix, Iterations: 3. Results were averaged.

But raw speed is just the beginning.

Hammerspace activates this local NVMe capacity by turning every GPU server into a node in its parallel global file system. It’s not just a mount—it’s a fully orchestrated performance tier, where files move automatically and intelligently between Tier 0 and any other storage tier or alternate region, based on business logic you define.

Need a dataset near the GPUs for training? Done. Need to move output to object storage for durability? Also done. No manual intervention. No forked copies. No siloed infrastructure.

This automated data orchestration is powered by Hammerspace policy-based Objectives, which operate at the file or object level, asynchronously and in real time. You get file granular control over data placement, protection, and performance—without ever slowing the GPUs down.

High-Performance File Storage - Same Performance as Lustre, ½ the Cost

In today’s hybrid, multi-cloud world, AI performance must scale without complexity, unnecessary overprovisioning, or runaway costs. Legacy file systems designed for HPC in the past are optimized to run in the datacenter, not in the cloud.

The chart below summarizes the costs of running a typical high performance storage environment in the cloud using Managed Lustre to the costs of running a similar environment in the cloud using Hammerspace. The costs of Root Disk and Archival Storage are shown for completeness, but costs have been kept the same for the purposes of this analysis.