NVIDIA、AI ネイティブストレージの暗号化・圧縮・整合性チェックを高速化

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. Encryption sits directly in the write path, where its throughput can determine how much secure data a platform can process each second before encryption becomes a bottleneck. Vera delivers up to 1.43x higher AES-128 encryption throughput than the x86 CPU used for comparison.

Decryption performs the corresponding operation on the read path andVera delivers up to 1.29x higher AES-128 decryption throughput than the x86 CPU used for comparison.

Higher encryption throughput enables storage systems to secure more data without constraining writes, while faster decryption reduces the time required to return protected data to agents, applications, or accelerators. These help storage systems protect and return growing volumes of AI data without consuming an increasing share of the storage-performance budget.

Protecting and recovering AI data faster

Storage platforms use erasure coding to protect data when drives, nodes, or data fragments become unavailable. Reed-Solomon encoding creates redundancy, while recovery uses that redundancy to reconstruct missing or corrupted data. Encoding typically occurs on the write path. Recovery occurs during rebuild, repair, or a degraded read. Their performance results can differ because these operations use different compute and memory-access patterns.

Vera delivers up to 3.26x higher Reed-Solomon throughput than the x86 CPU in a recovery workload.

Higher Reed-Solomon throughput enables storage systems to write protected data and reconstruct missing data faster. In selected efficiency measurements, Vera also completes more protection work within the available CPU power envelope, helping shorten rebuilds and reduce contention with normal data services.

Validating data integrity at higher throughput

Data must remain correct as it is moved, stored, and retrieved. Cyclic redundancy checks, or CRCs, create checksums that storage systems use to detect accidental corruption.

CRC and Reed-Solomon perform complementary roles. CRC detects that data no longer matches the expected result. Reed-Solomon provides the redundancy used to reconstruct missing or corrupted information.

A storage system may calculate CRCs on both the write and read paths, including while copying data between buffers. As data volumes grow, these checks can consume a meaningful share of CPU resources.

Vera delivers up to 3.67x higher CRC32C throughput than the x86 CPU.

Higher CRC32C throughput enables storage systems to validate more data without integrity checking limiting reads or writes. For agentic workloads, this helps return reliable context, persistent memory, and enterprise data with less CPU-side processing delay.

Reducing data footprint with faster compression and decompression

Agentic AI creates growing volumes of context, logs, checkpoints, retrieval data, intermediate outputs, and persistent memory. Compression reduces the storage capacity required for this data and the bandwidth needed to move it. Decompression restores the data when it is read. Compression may run inline or after data is written, depending on the storage architecture, while decompression is typically required when compressed data is read. Compression throughput affects how quickly data can be reduced, and decompression throughput can affect how quickly storage systems return data to agentic AI applications. The following benchmarks measure these operations independently.

Vera delivers up to 3.29x higher compression throughput than the x86 CPU, while sustaining its advantage across the measured thread counts.

The decompression benchmark measures the corresponding operation when compressed data is read. Vera delivers up to 1.72x higher decompression throughput than the x86 CPU under concurrency, with its advantage increasing as more worker threads run in parallel.

Compression performance varies by algorithm, data characteristics, compression level, buffer size, and thread count, so these results apply specifically to the measured workloads. Higher compression throughput enables storage systems to reduce the amount of data written, stored, and transferred while sustaining CPU-side processing performance.

Higher decompression throughput helps storage systems return decompressed data to agents and applications more quickly. Together, these capabilities lower pressure on storage capacity and bandwidth while enabling each processor to compress and decompress more data as agentic AI workloads grow.

Accelerating a multi-stage storage write path

Storage systems rarely execute data services independently. A secure write path may compress data to reduce its footprint and then encrypt it before it is written. The benchmark set includes a memory-resident pipeline that applies compression followed by encryption to each data buffer. Unlike the preceding benchmarks, which measure individual operations, this test measures total pipeline throughput when two CPU-intensive storage functions execute in sequence.

Vera delivers up to 3.21x higher pipeline throughput than the x86 CPU used in the benchmark for the two-stage compression and encryption pipeline.

This result shows that Vera’s performance advantage extends beyond the individual operations measured earlier to a multi-stage sequence representative of a storage write path that reduces and protects data.* *Higher multi-stage pipeline performance enables the storage system to process more data per processor, helping sustain write throughput as agentic AI data volumes grow.

Scaling agent execution and storage with Vera

Agentic AI makes CPU execution and storage processing part of the same AI factory data path.

The CPU runs tools, code, retrieval, analysis, and data-processing steps between model calls. The storage system secures, protects, validates, compresses, and returns the data those steps require. Both must scale within the finite CPU, power, and cooling resources available across the AI factory.

Vera was designed to accelerate CPU-dependent work for the agentic AI era. In NVIDIA Vera Rubin, it serves as the host CPU for NVIDIA GPUs and supports agent execution. Standalone Vera delivers up to 1.8x higher performance per core in agentic tools. In BlueField-4 STX, Vera powers the CPU-side processing used by AI-native storage platforms.

The benchmark results show Vera accelerating encryption and decryption for secure data access, Reed-Solomon recovery for faster reconstruction of missing or corrupted data during storage rebuilds and repairs, CRC32C for high-throughput integrity validation, and compression and decompression to reduce capacity and bandwidth demands and accelerate data retrieval.

By sustaining these functions individually and within a multi-stage write path, Vera helps AI-native storage platforms process and return secure, reliable data with less CPU-side delay. It also supports more concurrent data flows and higher service density without proportional growth in CPU resources, power, and cooling. Across selected workloads, Vera also delivers higher measured performance per watt, enabling more CPU-side storage processing within the available processor power budget.

A common Vera CPU architecture and software toolchain across compute and storage provides a consistent foundation for scaling both agent execution and the supporting data infrastructure.

Learn more about Vera CPUs and BlueField-4 STX

• Read NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI to learn more about the architecture powering these storage-processing results.

• Download the NVIDIA Vera BlueField-4 STX Storage Processor datasheet and learn how NVIDIA BlueField infrastructure processors accelerate storage, networking, and security for AI factories.