As AI Increases Demands on Memory, Storage Steps Up
| Source: NVIDIA Blog
Tags: NVIDIA, cuFile, GPUDirect Storage, Vera CPU, BlueField-4, AI infrastructure, storage
At the Future of Memory and Storage conference, NVIDIA announced its Vera CPU delivers 3.21x higher storage throughput than x86 for AI data pipelines, and open-sourced its cuFile GPUDirect Storage APIs—enabling GPUs to initiate storage reads directly without CPU mediation.
Details
At the Future of Memory and Storage (FMS) conference, NVIDIA presented storage infrastructure benchmarks and a key open-source release aimed at a growing AI workload bottleneck: thousands of concurrent AI agents demanding data faster than CPU-mediated storage can deliver. The Vera CPU, part of NVIDIA's BlueField-4 STX, benchmarks at 3.21x higher throughput than an x86 CPU in a two-stage compression and encryption pipeline. The argument is that as GPUs run thousands of concurrent storage requests, the CPU becomes the bottleneck for the encryption, compression, and verification operations those requests require—particularly relevant as AI agent deployments scale. NVIDIA is open-sourcing its cuFile APIs, the layer that enables GPUs to read from and write to storage directly using hundreds of thousands of GPU threads. This bypasses the CPU for data path operations and brings access latency down to microseconds. cuFile is a component of NVIDIA GPUDirect Storage and its open sourcing increases interoperability between GPU vendors and storage manufacturers.