NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage

NVIDIA has released benchmark data for its Vera BlueField-4 STX Storage Processor, demonstrating marked performance advantages over commodity x86 CPUs in essential storage operations. This development directly addresses a growing bottleneck in AI infrastructure, where the CPU-intensive tasks of securing, protecting, and preparing data are struggling to keep pace with the demands of concurrent, large-scale agentic AI workflows that continuously access and generate vast amounts of information.

The new benchmarks isolate foundational storage primitives to measure the processor's direct contribution to throughput. The Vera processor integrates 88 Olympus Armv9.2 cores, Spatial Multithreading, and a high-bandwidth SOCAMM2 LPDDR5X memory subsystem to accelerate these tasks without relying on the host system's CPU. By offloading this work, Vera enables storage platforms to process more data while applying enterprise-grade services more efficiently.

Vera Benchmark Performance vs. x86

Key performance gains highlighted in the report include:

• Reed-Solomon recovery: Up to 3.26x faster
• CRC32C integrity checking: Up to 3.67x faster
• Compression (Zstandard/LZ4): Up to 3.29x faster
• AES-128 Encryption/Decryption: Up to 1.43x and 1.29x faster, respectively
• Multi-stage pipeline operations: Up to 3.21x faster

The performance uplift has significant implications for the broader AI market. By moving intensive storage processing to a specialized unit like the Vera BlueField-4 STX, data centers can achieve higher service density and support more concurrent AI agents within smaller power and cooling envelopes. This architectural approach helps ensure expensive GPU resources remain fully utilized, preventing storage operations from becoming the limiting factor in application responsiveness and token generation for complex AI systems.

Strategic Takeaway: NVIDIA's benchmark release for Vera reinforces an architectural shift in the AI data center, where data-centric tasks like encryption and compression are moved from general-purpose CPUs to specialized processors to eliminate I/O bottlenecks and maximize the efficiency of the entire accelerated computing stack.