Too slow

Optimised AI projects require storage with ultra-low latency and incredible bandwidth. Performance previously only found on enterprise class storage solutions.

Too large

Legacy enterprise storage is designed to scale to many petabytes and serve 100’s of clients, well in excess of the majority of AI projects.

Too expensive

As a result, traditional enterprise storage is overengineered and overpriced for 90% of AI projects which run on relatively small GPU clusters.

The AI Storage Challenge

AI application performance is determined by three major elements: GPU compute, data storage and the network. As GPU compute performance continues to rise at an astounding rate they consume and analyse data at much higher rates than many storage solutions can deliver, resulting in low utilisation of expensive GPU resources and dramatically extended training and project times.

Legacy storage solutions have evolved for traditional IT applications, rich in enterprise features unnecessary for AI. Parallel filesystems are designed for HPC environments. Both options are complex and expensive, designed for much larger solutions and impact the budget for GPUs, draining resources away from innovation.

See our partner's Scan explainer video

The PEAK:AIO Solution

PEAK:AIO is a Software Defined Storage platform designed from the ground up to meet the specific needs of AI Applications. Deployed on server technology from the likes of Dell to create an AI Data Server that delivers the ultra-low latency and incredible bandwidth needed to optimise GPU utilisation.

Ideally suited to mid-scale GPU clusters where the traditional storage route is to complex and expensive.

GPU-ready performance

Using GPUDirect and RDMA protocols PEAK:AIO AI data servers are capable of delivering over 40GB/sec per 2U to the GPU compute, outperforming some of the largest multi-node solutions, in a simple cost-effective 2U footprint.

Evolved file system and features

Maintains the simplicity of the network filesystem (NFS) with turbo charged wire-speed performance. Coupled with ultra-low latency NVMe-oF, the AI Data Server is simple to use with no complex drivers or configuration and provides breakthrough performance with zero administration.

Fault Tolerance and Flexible Features

Working with the large AI userbase PEAK:AIO has incorporated and tuned the features relevant to the typical AI workflow and removed legacy storage features which introduce latency. PEAK:AIO includes enhanced performance RAID protection which produce the performance and world leading kernel compliant connectivity options and protocols which deliver the performance. Advance technology, yet simple to use.

Scalable

AI projects evolve, PEAK:AIO allows organisations to start as small as 30TB and grow seamlessly up to 360TB per 2U as projects ramp. No need to overinvest in storage at the begining, simply grow as needed by adding additional drives or 2U Data Servers

Cost effective

PEAK:AIO dramatically reduces the cost of high-performance AI data storage, enabling a larger percentage of budget to be spent on GPU compute. Providing a more effective overall solution, faster project results and greater return on investment.

Typical Configuration

This diagram illustrates a typical AI cluster using a Dell AI Data Server powered by PEAK:AIO. Often early concept projects use only one server, PEAK:AIO can work directly connected, removing the cost of the switch at this initial stage. As the compute scales, the simple addition of a switch enables the PEAK:AIO solution to serve up to 10 GPU servers such as the NVIDIA DGX A100’s shown below.