NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage

來源原標題

NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage

作者

Elizabeth Goodman

原始標籤

Agentic AI / Generative AI · Data Center / Cloud · Networking / Communications · AI Agent · AI Factory · BlueField DPU · Cloud APIs · Cloud Networking · Cloud Services · DOCA · Software-Defined Data Center · Vera CPU · Vera Rubin · Vera Rubin NVL72

原始連結

https://developer.nvidia.com/blog/nvidia-vera-storage-benchmarks-faster-encryption-compression-integrity-checking-and-recovery-for-ai-native-storage/

摘要原文

Storage is an active part of every agentic AI workflow. As agents retrieve enterprise knowledge, access persistent memory, reuse key-value (KV) cache data, execute tools, and generate new results, storage systems must continuously supply and preserve the data that moves the agent reasoning loop. Each agent step can trigger multiple storage operations, and those operations can repeat across thousands of concurrent agents with increasingly larger context windows. Supplying and preserving this data requires more than basic reads and writes. AI inference runs on GPUs, but agentic processes, tool calls, data management tasks, and the storage services that support them run on CPUs. During writes, storage may compress and encrypt data, calculate checksums, and calculate redundancy. During reads, it may validate, decrypt, decompress, or reconstruct data before returning it to the application. These functions are essential to the security and resilience of AI systems. Each function also requires additional CPU processing as data moves through the storage path. As agent concurrency (multiple users, AI agents, or tasks running in parallel) and context volumes grow, storage must perform more of this work without constraining application responsiveness or token generation; it must supply data at the rate required for accelerated computing. Many of these functions sit directly in the data path; one delayed operation can slow the broader data flow. Scaling them with conventional CPUs can require more cores, power, and cooling, increasing infrastructure cost while still leaving performance dependent on the slowest step. Faster SSDs and networks cannot deliver their full potential if the processor securing, protecting, and preparing the data cannot keep pace. Closing the storage processing gap The NVIDIA Vera BlueField-4 STX Storage Processor , a key component of the NVIDIA STX foundation for AI-native data platforms, brings the NVIDIA Vera CPU performance directly into the storage data path. The same Vera CPU architecture designed to keep NVIDIA Rubin GPUs fed also accelerates the CPU-side storage processing. The benchmark results show Vera outperforming the x86 CPU across encryption and decryption, recovery, integrity checking, compression and decompression, and a multi-stage storage pipeline. These gains enable storage platforms to process more data and apply essential enterprise services with less CPU and power overhead, while higher compression throughput helps reduce storage capacity and bandwidth demands. This post explains how the Vera CPU in BlueField-4 STX accelerates the storage processing required by agentic AI, helping AI-native storage platforms secure, protect, validate, and compress more data while increasing storage-processing throughput and efficiency. Vera CPU architecture: Built for storage’s dual demands Vera CPU includes 88 NVIDIA-designed Olympus CPU cores that are fully compatible with the Armv9.2 instruction set. The CPU supports 176 NVIDIA Spatial Multithreading threads. It pairs these cores with the NVIDIA Scalable Coherency Fabric (SCF), Small Outline Compression Attached Memory Module (SOCAMM2) LPDDR5X memory to sustain strong single-thread performance, and high-throughput CPU execution at AI-factory scale. The SCF provides a coherent, on-die data path across the cores, shared cache, memory controllers, and I/O, with up to 3.4 TB/s of bisection bandwidth and a 164 MB unified L3 cache. This gives active cores high-bandwidth, predictable access to shared data as workloads scale across the processor. The SOCAMM2 LPDDR5X memory subsystem complements the fabric with up to 1.2 TB/s of aggregate memory bandwidth, or up to 14 GB/s per core, helping keep the NVIDIA Olympus cores supplied across bandwidth-intensive and highly concurrent workloads. Modular, field-replaceable memory combines LPDDR5X power efficiency with the serviceability and reliability required for datacenter infrastructure. Storage primitives place two distinct demands on a CPU. First, within each data stream, encryption, integrity checking, recovery, compression, and decompression must complete quickly before subsequent storage processing can proceed, making sustained per-core performance important. Second, across the system, these operations run over many concurrent streams and repeatedly move data through caches and memory, making bandwidth and predictable latency equally important. Vera addresses both requirements. The Olympus core combines wide instruction throughput, advanced branch prediction, deep out-of-order execution, and vector and cryptographic resources to help each core sustain instruction throughput across control-heavy and data-processing code. NVIDIA Spatial Multithreading, the monolithic compute die, SCF, unified L3 cache, and high-bandwidth SOCAMM2 memory help keep active cores supplied with data while reducing thread-to-thread interference and supporting more predictable data access under load. Together, these capabilities help explain the measured gains across encryption, integrity checking, parity calculations, compression, and the multi-stage storage pipeline. This enables the BlueField-4 STX Storage Processor to sustain more CPU-side storage processing across concurrent data streams without proportional increases in CPU resources, power, and cooling. Measuring foundational storage performance Storage tasks execute repeatedly across storage read, write, and recovery paths. Their throughput and efficiency help determine whether CPU-side processing keeps pace with SSDs and networks or becomes the limiting stage in the data path. The storage primitive microbenchmarks in this post isolate these functions to measure the processor’s contribution. They show the CPU performance and headroom available in Vera for building higher-throughput, more efficient storage services. Each test runs within a single process using data already held in memory. The tests exclude file I/O, disk performance, networking, command startup, and external-device bottlenecks unless otherwise identified. The benchmark set uses common libraries, including OpenSSL, Zstandard, and LZ4, along with comparable implementations optimized to use the native instructions available on Arm and x86 processors. A purpose-built test framework runs each workload consistently and controls buffer sizes, thread counts, CPU placement, timing, correctness validation, and result collection. The algorithms and many of the software implementations are widely adopted. The test framework applies the same workload definitions and controls across Vera and x86, enabling a consistent processor comparison. The results can be reproduced using the source code, scripts, fixed software versions, configurations, and result files, although the complete benchmark set is not an off-the-shelf public benchmark. These measurements establish Vera’s performance on the storage building blocks that influence secure data movement, resilience, capacity efficiency, and service density. Production storage paths often apply several of these operations to the same data, causing their CPU-processing requirements to accumulate. When combined in production storage software, the performance of these individual operations contributes to aggregate throughput and CPU efficiency. Higher throughput across individual primitives and the multi-stage pipeline gives storage software more CPU headroom to keep pace with SSDs and networks, support concurrent data flows, and apply essential data services efficiently. End-to-end testing is still required to quantify complete storage-system or GPU-performance outcomes. Securing more AI data with faster encryption AI factories process sensitive information, including model assets, enterprise knowledge, agent context, prompts, outputs, and customer data. AES-128 is widely used for data-at-rest and data-in-flight encryption.