Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage

NVIDIA today announced benchmark results for its Vera BlueField-4 STX Storage Processor, demonstrating significant performance gains over traditional x86 CPUs for encryption, compression, integrity checking, and data recovery in AI-native storage systems. The processor, built around the NVIDIA Vera CPU architecture, is designed to accelerate the storage data path that underpins agentic AI workflows, where thousands of concurrent agents repeatedly read and write knowledge, memory, and key-value cache data. According to NVIDIA, the Vera CPU delivers up to 2.5x faster encryption and decryption, 3x faster compression, and 2x faster integrity checks compared to a leading x86 server CPU, while consuming 40% less power per operation. These benchmarks come as storage becomes an active bottleneck in AI factories, where GPUs demand data faster than conventional CPUs can secure and validate it.

The announcement addresses a critical bottleneck in modern AI infrastructure. Agentic AI workflows—where AI agents retrieve enterprise knowledge, access persistent memory, and execute tools—generate a flurry of storage operations. Each agent step can trigger multiple reads and writes, multiplied across thousands of concurrent agents with growing context windows. Storage must compress, encrypt, checksum, and reconstruct data on the fly, tasks that traditionally fall to the CPU and can severely lag behind GPU processing speeds. NVIDIA argues that faster SSDs and networks cannot reach their full potential if the processor securing the data cannot keep pace, and its new benchmarks show Vera closing this gap.

Vera delivers its performance through a purpose-built architecture. The processor integrates 88 NVIDIA-designed Olympus CPU cores, fully compatible with the Armv9.2 instruction set, and supports 176 NVIDIA Spatial Multithreading threads. It pairs these cores with the NVIDIA Scalable Coherency Fabric (SCF), which provides a coherent, on-die data path with up to 3.4 TB/s of bisection bandwidth and a 164 MB unified L3 cache. The system also uses SOCAMM2 LPDDR5X memory modules, offering up to 1.2 TB/s of aggregate memory bandwidth—or 14 GB/s per core—to sustain bandwidth-intensive storage workloads. This combination allows active cores to access shared data with predictable, high-speed performance as workloads scale.

The practical impact is substantial for storage platforms serving AI workloads. By offloading encryption, compression, and integrity checking to Vera, storage systems can process far more data with fewer CPU cores and lower power consumption. Higher compression throughput, for example, directly reduces storage capacity and network bandwidth requirements, while faster recovery ensures quicker response in fault-tolerant systems. Enterprise storage vendors adopting AI-native platforms can maintain application responsiveness and token generation rates without overprovisioning conventional server CPUs.

Industry analysts have responded cautiously, noting that while NVIDIA's benchmarks are compelling, real-world performance will depend on integration and software ecosystem maturity. The shift toward Arm-based storage processors signals a broader trend of specialized silicon entering the storage path, following similar moves in networking and security. NVIDIA’s focus on storage is a natural extension of its AI platform dominance, as the company seeks to control the entire data pipeline from GPU to storage.

Moving forward, NVIDIA plans to fold Vera into the full BlueField-4 STX foundation, enabling storage platforms to apply advanced data services at AI-factory scale. The company expects early adopters to test the processors in production environments later this year, with broader availability in 2026. This development accelerates the transition to AI-native data platforms, where storage is no longer a passive repository but an active, accelerated participant in the AI reasoning loop.

More broadly, Vera represents an acknowledgment that AI performance is not solely a GPU story. As inference and training demand ever-increasing data throughput, CPUs must evolve to handle data movement and protection efficiently. NVIDIA’s benchmarks signal a future where every element of the AI stack—compute, network, and storage—benefits from purpose-built, high-efficiency processor architectures. For enterprises deploying AI agents at scale, this could mean lower infrastructure costs, faster responses, and more resilient data pipelines in the race to production AI.

Turn any website into clean JSON.

Ready-made scrapers & AI agents

Try Apify