NVIDIA Releases Vera BlueField-4 STX Storage Processor Benchmark
NVIDIA has announced the latest benchmark for the Vera BlueField-4 STX storage processor, showing that the AI-native storage platform based on the Vera CPU significantly outperforms traditional x86 CPUs in tasks such as encryption and compression.
The tests indicate that AES-128 encryption throughput is improved by up to 1.43 times, Reed-Solomon data recovery by up to 3.26 times, CRC32C checks by up to 3.67 times, and compression throughput by up to 3.29 times, with an overall maximum improvement of 3.21 times across multi-stage processes; the Vera CPU features an 88-core Olympus architecture, supporting 176 threads and interconnect bandwidth of up to 3.4TB/s.
Funding for AI factories and storage system procurement may accelerate towards integrated Vera solutions, benefiting NVIDIA's ecosystem partners while putting pressure on storage and CPU suppliers reliant on traditional x86.
Source: Public Information
ABAB AI Insight
NVIDIA has been continuously developing its own Arm architecture cores since the Grace CPU, with Vera further enhancing single-thread and full-load performance using the Olympus core. Previously, BlueField-4 STX was announced at GTC as an AI-native storage reference architecture, indicating a historical trend towards unifying CPU, DPU, and GPU into a collaborative AI factory system.
Capital is shifting from general-purpose server CPUs to AI-specific data processing units, motivated by the higher throughput demands of KV caching, tool invocation, and long-term memory as AI agents scale. This is reflected in embedding the Vera CPU into storage data paths, reducing host CPU usage and improving efficiency in critical tasks like encryption and compression.
Similar cases can be seen in the early evolution of DPUs offloading networking and security. The current AI infrastructure landscape is transitioning from GPU-centric to a full-stack collaboration of compute-storage-network, with control shifting towards chip manufacturers that can provide end-to-end acceleration.
This essentially represents a technological replacement: the bottleneck of traditional x86 in storage data paths is being replaced by the dedicated Vera architecture, which offers higher bandwidth interconnects and core designs optimized for AI workloads, thus synchronously enhancing the efficiency of storage processing and inference.
ABAB News · Cognitive Laws
- Storage bottlenecks determine the limits of AI factories.
- Dedicated architectures consume data faster than general cores.
- Offloading paths represent the new frontier of computation.