VAST Lightspeed - Storage for the AI Decade

Companies, research organisation and government are all focused on implementing artificial intelligence and machine learning to assist them to deliver better services, better customer outcomes and more cost efficient operations. XENON has been involved at the forefront of AI implementations in Australia for over 10 years, and in that time we seen the processing power of GPUs driving AI increase exponentially while the sophistication of AI models multiplies. The foundation of the growth in AI is the exponential growth in data, the raw material for AI workloads. Fundamentally, the better the data, the better the AI models – from initial analytics, to model training, to inference.

Implementing or scaling your AI will challenge your data storage strategies:

Do you keep AI data combined with all the business data?
If business data is your AI data, how do you use it for both business and AI without impacting access for each?
At petabyte scale, how do you manage costs and performance?
AI models perform best when all model data is instantly available, so how is this managed across storage infrastructure that is typically split into high performance, nearline archive and cold storage.

VAST Lightspeed – New Storage Technology and a New Approach to Storage

VAST Data solves these problems with a fundamentally new storage offering which provides the speed of flash, the simplicity of NAS, and exabyte scale and cost efficiency. VAST achieves this by combining Intel Optane SSDs, QLC Flash, and NVMe-over-fabrics networking to create a disaggregated, shared everything storage architecture.

AI Demands Fast Access to All your Data, All the Time

For AI, this is truly revolutionary. AI workloads read and write data in manner unlike any other applications. Models access data in a random fashion, based on the inputs they are facing at the time. The effect is random read-writes across the entire library of AI data, all the time. This makes tiering storage strategies untenable, as pushing data to slower, cheaper tiers will negatively impact performance when that data is required.

AI compute speeds, such as in the DGX A100 now reach 200 GB/s and the challenge is to create a storage architecture to meet these demands. Typically, the limiting factor has been the CPU as all GPU data requests are routed through the CPU, to the storage, and back again. Now, NVIDIA’s GPUDirect Technology, allows NVIDIA GPUs to access storage directly, eliminating buffer bounce and streamlining GPU access to data.

Now, the critical question really is how fast can your storage architecture feed the AI engine. With VAST Lightspeed storage, you can now deliver the storage performance your AI infrastructure demands. VAST has benchmarked Lightspeed performance with the NVIDIA DGX A100 – the fastest AI data centre servers.

In addition to being blindingly fast, VAST Lightspeed with GPUDirect also requires far fewer cores to deliver this performance – further accelerating your workloads by freeing up your compute resources.