Everpure has set out an AI data architecture intended to stop costly GPU servers sitting idle while they wait for data, in a sponsored feature published by The Register.
The vendor's VP of AI Infrastructure, Par Botes, said the characteristics of performance, resilience, availability and governance do not fundamentally change in the AI era, but access patterns are different. He said metadata can become richer than the data itself because it explains the semantics of data and how it behaves as it becomes enriched.
Everpure identifies three bottlenecks. The first is throughput starvation at scale: keeping thousands of GPUs saturated requires storage bandwidth to scale past 4 to 5 TB/sec to 10 TB/sec or more, which it says rules out standard dual-controller arrays and siloed direct-attached storage. Its FlashBlade//S and FlashBlade//EXA systems are presented as platforms built for AI and HPC, with FlashBlade//EXA described as Nvidia AI Data Platform-certified and delivering 400 to 450 million IOPS, 220 GB/sec bandwidth and 4.6 billion metadata operations/sec.
The second is what Everpure calls the KV cache prefill tax. When a model finishes a run, tokens held in GPU high-bandwidth memory are discarded, so a later query against the same document can require the prefill to be recomputed. Everpure Key Value Accelerator offloads cached token states to shared flash over NVIDIA GPUDirect Storage via RDMA, so pre-computed KV tensors can stream into GPU memory and bypass host CPU overhead.
Botes said many patterns are far more reusable than people think, and that the more they are reused, the more compute time can be saved. He also said data speed drops dramatically as it moves further from the CPU or GPU, describing nanoseconds for local memory, microseconds for disk and milliseconds for network transfers.
The third bottleneck is data silos. Everpure argues that copying data at scale is becoming impossible, and that once source data changes, a data lake becomes invalid. Its Data Stream platform is described as ingesting, curating and transforming data, vectorising and indexing it, with built-in vector databases using Nvidia's NeMo Retriever for GPU-accelerated vector embedding generation in RAG pipelines.