NVIDIA Launches AI Storage Accelerator: Data Encryption, Compression, and Recovery Speeds Boosted by Up to 3.67x

August 4th, NVIDIA announced the latest benchmark test results of the Vera BlueField-4 STX storage processor, showing that its AI-native storage platform based on the Vera CPU architecture achieved significant performance improvements in encryption, compression, data integrity verification, and recovery compared to traditional x86 CPUs.

NVIDIA stated that as AI Agent applications scale up, storage systems need to continuously process enterprise knowledge bases, long-term memory, KV caches, tool invocation data, and model generation results, with traditional CPUs gradually becoming a performance bottleneck in the data processing path.

The tests showed that the Vera CPU outperformed the compared x86 CPU in multiple storage tasks:

AES-128 encryption throughput increased by up to 1.43 times, decryption throughput increased by up to 1.29 times;

Reed-Solomon data recovery performance increased by up to 3.26 times;

CRC32C data integrity verification performance increased by up to 3.67 times;

Compression throughput increased by up to 3.29 times, decompression performance increased by up to 1.72 times;

In the multi-stage storage process of compression + encryption, overall throughput increased by up to 3.21 times.

The Vera CPU adopts NVIDIA's independently designed Olympus core architecture, which includes 88 Armv9.2 compatible CPU cores, supports 176 threads, and is equipped with Scalable Coherency Fabric (SCF) and SOCAMM2 LPDDR5X memory system, providing a maximum of 3.4TB/s interconnect bandwidth and a maximum of 1.2TB/s memory bandwidth.

NVIDIA stated that the AI factory not only relies on GPUs for model inference, but also requires high-performance CPUs and storage infrastructure support when AI Agents perform tool invocation, data retrieval, and task processing. By introducing Vera CPU capability into the storage data path, BlueField-4 STX can help the AI-native storage platform reduce CPU resource, power, and heat pressures.

NVIDIA said that future AI workloads will bring higher concurrency, larger context scales, and more data processing requirements, and the Vera architecture aims to enhance the collaborative efficiency of computing, storage, and AI inference infrastructure in data centers.