NVIDIA Vera Storage Benchmarks Show Faster Encryption, Compression, and Recovery

NVIDIA published benchmarks on August 3, 2026, showing that its Vera CPU, embedded in the BlueField-4 STX storage processor, delivers higher throughput than a benchmarked x86 CPU for encryption, compression, integrity checking, and Reed-Solomon recovery. The tests, authored by Jason Hardy and Ronil Prasad, use OpenSSL, Zstandard, and LZ4 libraries optimized for native Arm instructions. Each test runs in a single process with data held in memory, excluding file I/O and networking bottlenecks. The framework controls buffer sizes, thread counts, CPU placement, timing, and correctness validation, and results are reproducible with provided source code, scripts, fixed software versions, and configurations.

Hardware and Test Setup

The Vera CPU contains 88 Olympus Armv9.2 cores and 176 spatial-multithreading threads. Its SCF provides up to 3.4 TB/s bisection bandwidth and 164 MB unified L3 cache. SOCAMM2 LPDDR5X memory offers up to 1.2 TB/s aggregate bandwidth, about 14 GB/s per core. The Vera BlueField-4 STX storage processor embeds the Vera CPU in the storage data path, meaning storage operations can run on the same processor that handles the data flow.

Benchmark Results

NVIDIA reports the following throughput improvements over the benchmarked x86 CPU: AES-128 encryption up to 1.43 times higher, AES-128 decryption up to 1.29 times higher, Reed-Solomon recovery up to 3.26 times higher, CRC32C integrity checking up to 3.67 times higher, compression up to 3.29 times higher, and decompression up to 1.72 times higher. A two-stage compression-and-encryption pipeline runs up to 3.21 times faster on Vera. These figures are vendor claims, not independent measurements.

Implications for Storage Workloads

For engineers serving models, the practical takeaway is that storage processing tasks, which often compete with inference workloads for CPU cycles, can be offloaded to a dedicated processor. The Vera architecture reduces the required CPU cores, power, and cooling for AI-native storage workloads, according to NVIDIA. This could free up x86 cores for inference or other compute-intensive tasks. The benchmarks focus on CPU-side storage operations, not on GPU or network performance. NVIDIA suggests readers consult the Vera CPU article and the BlueField-4 STX datasheet for more details.