NVIDIA has published storage benchmark results for its Vera CPU, saying the NVIDIA BlueField-4 STX storage processor can improve encryption, compression, integrity verification, and data recovery performance in AI-native storage environments.
The company said the results are aimed at enterprise data demands in the age of agentic AI. As AI agents perform knowledge retrieval, call tools, manage long-term memory, and work with larger context windows, storage systems are no longer just read-and-write components. NVIDIA described storage as a key part of the AI inference path.
That shift puts more work on the storage path itself. Large volumes of data need to be encrypted, compressed, checked, and recovered, tasks that are typically handled by CPUs and can become a performance bottleneck for AI infrastructure.
Benchmark results against a comparison x86 CPU
NVIDIA said the BlueField-4 STX storage processor with Vera CPU showed higher performance than a comparison x86 CPU across multiple storage workloads:
- AES-128 encryption: up to 1.43x
- AES-128 decryption: up to 1.29x
- Reed-Solomon data recovery: up to 3.26x
- CRC32C integrity checking: up to 3.67x
- Compression: up to 3.29x
- Decompression: up to 1.72x
- Multi-stage compression-plus-encryption storage pipeline: up to 3.21x
NVIDIA says AI-native storage needs low latency at higher scale
NVIDIA said scaling storage processing with traditional CPUs usually requires adding more cores, which also raises power consumption and cooling costs. AI-native storage, by contrast, needs to maintain low latency under heavier concurrency and larger data volumes.
According to NVIDIA, Vera improves the amount of work handled per unit of CPU resource, allowing storage systems to support more AI agent workloads without a significant increase in infrastructure cost.