NVIDIA on Aug. 4 published new benchmark results for its Vera BlueField-4 STX storage processor, saying its Vera CPU-based AI-native storage platform delivered higher performance than traditional x86 CPUs in core tasks including encryption, compression, data integrity checks, and recovery.
The company said storage systems are under rising pressure as AI agent applications scale. Those systems need to keep processing enterprise knowledge bases, long-term memory, KV cache, tool-call data, and model-generated outputs. In NVIDIA’s description, the CPU is becoming a bottleneck in the data path.
Benchmark results across storage workloads
NVIDIA said Vera CPU outperformed the compared x86 CPU in several storage tasks.
- AES-128 encryption throughput improved by as much as 1.43x, while decryption throughput improved by up to 1.29x.
- Reed-Solomon data recovery performance increased by as much as 3.26x.
- CRC32C data integrity check performance rose by as much as 3.67x.
- Compression throughput improved by up to 3.29x, and decompression performance by as much as 1.72x.
- In a multi-stage storage flow combining compression and encryption, overall throughput improved by as much as 3.21x.
Vera CPU specifications
According to NVIDIA, Vera CPU is built on the company’s in-house Olympus core architecture. It includes 88 Armv9.2-compatible CPU cores and supports 176 threads. The chip also uses Scalable Coherency Fabric, or SCF, and a SOCAMM2 LPDDR5X memory system.
NVIDIA said the platform can provide up to 3.4 TB/s of interconnect bandwidth and up to 1.2 TB/s of memory bandwidth.
Focus on AI agent data paths
The company said AI factories do not rely only on GPUs for model inference. When AI agents execute tool calls, retrieve data, and handle tasks, they also require high-performance CPUs and storage infrastructure. NVIDIA said BlueField-4 STX brings Vera CPU capabilities into the storage data path, which can help AI-native storage platforms reduce CPU resource use as well as power and cooling pressure.
NVIDIA added that future AI workloads will bring higher concurrency, larger context sizes, and greater data processing demand. Vera is designed to improve coordination across compute, storage, and AI inference infrastructure in data centers.