NVIDIA Vera BlueField-4 STX Benchmarks Show Storage Acceleration for AI-Native Platforms

NVIDIA published benchmark results demonstrating that its Vera CPU, embedded in BlueField-4 STX storage processors, outperforms a comparable x86 CPU across key storage primitives—including encryption, Reed-Solomon recovery, CRC32C integrity checking, compression, decompression, and a multi-stage write path—by factors up to 3.67x. These results position Vera as a CPU-side accelerator for AI-native storage, addressing growing storage processing demands from agentic AI workloads.

What happened

Source factNVIDIA published benchmark results for the Vera CPU in the BlueField-4 STX Storage Processor, comparing its performance against an x86 CPU on storage primitives such as AES-128 encryption/decryption, Reed-Solomon recovery, CRC32C, compression, decompression, and a two-stage compress-encrypt pipeline.

Source factVera CPU includes 88 NVIDIA-designed Olympus cores compatible with Armv9.2, 176 Spatial Multithreading threads, a Scalable Coherency Fabric with 3.4 TB/s bisection bandwidth, a 164 MB unified L3 cache, and SOCAMM2 LPDDR5X memory with up to 1.2 TB/s aggregate bandwidth.

What is actually new

Source factVera delivered up to 1.43x encryption, 1.29x decryption, 3.26x Reed-Solomon recovery, 3.67x CRC32C, 3.29x compression, 1.72x decompression under concurrency, and 3.21x pipeline throughput versus the x86 CPU in the benchmark set.

Why it matters

AI analysisThe CPU-side storage functions are directly in the data path, and delays can slow the broader data flow. Vera's throughput gains suggest that storage systems can maintain secure, protected, and validated data delivery with less CPU overhead, which is crucial as agent concurrency and context volumes increase.

Constraint shift

AI analysisBy demonstrating higher throughput and possible power efficiency across the tested storage primitives, Vera shifts the limiting resource away from CPU processing, enabling storage platforms to exploit faster SSDs and network devices more fully.

Implications

AI analysisIf independently verified, these benchmarks imply that AI-native storage architecture should treat CPU choice as a strategic decision. NVIDIA can leverage Vera across compute and storage to offer a consistent platform for AI factories, potentially displacing traditional x86-based storage controllers.

What would change my mind

AI hypothesisMy confidence would decrease if independent tests, using identical or more realistic end-to-end storage workloads, fail to reproduce the expected gains, or if actual system-level throughput is dominated by network or SSD bottlenecks. Power efficiency claims require detailed measurement to be fully evaluated.