NVIDIA's Vera processor beats x86 CPUs in benchmarks

NVIDIA has unveiled benchmark results for its Vera BlueField-4 STX Storage Processor, demonstrating substantial performance gains over traditional x86 CPUs for AI-native data platforms. The hardware integrates 88 NVIDIA-designed Olympus Armv9.2 cores supporting 176 Spatial Multithreading threads. It features a Scalable Coherency Fabric that delivers up to 3.4 TB/s of bisection bandwidth and a 164 MB unified L3 cache, paired with a SOCAMM2 LPDDR5X memory subsystem providing up to 1.2 TB/s of aggregate memory bandwidth, or up to 14 GB/s per core. This architecture is designed to supply data to NVIDIA Rubin GPUs by accelerating CPU-side storage processing directly in the data path.

In isolated microbenchmarks comparing the processor to an x86 CPU, the Vera architecture achieved up to 1.43x higher throughput for AES-128 encryption and up to 1.29x for decryption. For data protection and validation, it reached up to 3.26x higher throughput in Reed-Solomon recovery workloads and up to 3.67x higher throughput for CRC32C integrity checking. Storage efficiency tests showed up to 3.29x higher compression throughput and up to 1.72x higher decompression throughput. Furthermore, in a multi-stage pipeline executing sequential compression and encryption, Vera delivered up to 3.21x higher throughput. Standalone Vera also achieved up to 1.8x higher performance per core in agentic tools.

For infrastructure practitioners and system architects, these performance leaps address the critical bottlenecks of agentic AI workflows. Because agentic processes repeatedly trigger complex storage operations across thousands of concurrent agents with expanding context windows, traditional x86 CPUs often struggle to keep pace without massive power and cooling overhead. By offloading and accelerating these foundational storage tasks, the BlueField-4 STX allows practitioners to scale concurrent data flows and increase service density. This ensures that security, integrity checking, and data reduction do not delay the reasoning loops of accelerated GPU systems, ultimately lowering the total cost and physical footprint of AI factories.