Storage is an energetic a part of each agentic AI workflow. As brokers retrieve enterprise data, entry persistent reminiscence, reuse key-value (KV) cache knowledge, execute instruments, and generate new outcomes, storage programs should constantly provide and protect the information that strikes the agent reasoning loop.
Every agent step can set off a number of storage operations, and people operations can repeat throughout hundreds of concurrent brokers with more and more bigger context home windows. Supplying and preserving this knowledge requires greater than fundamental reads and writes. AI inference runs on GPUs, however agentic processes, device calls, knowledge administration duties, and the storage companies that assist them run on CPUs.
Throughout writes, storage might compress and encrypt knowledge, calculate checksums, and calculate redundancy. Throughout reads, it might validate, decrypt, decompress, or reconstruct knowledge earlier than returning it to the applying. These features are important to the safety and resilience of AI programs. Every operate additionally requires extra CPU processing as knowledge strikes by means of the storage path.
As agent concurrency (a number of customers, AI brokers, or duties operating in parallel) and context volumes develop, storage should carry out extra of this work with out constraining utility responsiveness or token technology; it should provide knowledge on the price required for accelerated computing.
Many of those features sit immediately within the knowledge path; one delayed operation can gradual the broader knowledge circulate. Scaling them with standard CPUs can require extra cores, energy, and cooling, growing infrastructure price whereas nonetheless leaving efficiency depending on the slowest step. Sooner SSDs and networks can not ship their full potential if the processor securing, defending, and getting ready the information can not hold tempo.
Closing the storage processing hole
The NVIDIA Vera BlueField-4 STX Storage Processor, a key element of the NVIDIA STX basis for AI-native knowledge platforms, brings the NVIDIA Vera CPU efficiency immediately into the storage knowledge path. The identical Vera CPU structure designed to maintain NVIDIA Rubin GPUs fed additionally accelerates the CPU-side storage processing.
The benchmark outcomes present Vera outperforming the x86 CPU throughout encryption and decryption, restoration, integrity checking, compression and decompression, and a multi-stage storage pipeline. These positive factors allow storage platforms to course of extra knowledge and apply important enterprise companies with much less CPU and energy overhead, whereas greater compression throughput helps scale back storage capability and bandwidth calls for.
This submit explains how the Vera CPU in BlueField-4 STX accelerates the storage processing required by agentic AI, serving to AI-native storage platforms safe, defend, validate, and compress extra knowledge whereas growing storage-processing throughput and effectivity.
Vera CPU structure: Constructed for storage’s twin calls for
Vera CPU consists of 88 NVIDIA-designed Olympus CPU cores which can be absolutely appropriate with the Armv9.2 instruction set. The CPU helps 176 NVIDIA Spatial Multithreading threads. It pairs these cores with the NVIDIA Scalable Coherency Cloth (SCF), Small Define Compression Hooked up Reminiscence Module (SOCAMM2) LPDDR5X reminiscence to maintain robust single-thread efficiency, and high-throughput CPU execution at AI-factory scale.
The SCF supplies a coherent, on-die knowledge path throughout the cores, shared cache, reminiscence controllers, and I/O, with as much as 3.4 TB/s of bisection bandwidth and a 164 MB unified L3 cache. This provides energetic cores high-bandwidth, predictable entry to shared knowledge as workloads scale throughout the processor. The SOCAMM2 LPDDR5X reminiscence subsystem enhances the material with as much as 1.2 TB/s of combination reminiscence bandwidth, or as much as 14 GB/s per core, serving to hold the NVIDIA Olympus cores provided throughout bandwidth-intensive and extremely concurrent workloads. Modular, field-replaceable reminiscence combines LPDDR5X energy effectivity with the serviceability and reliability required for datacenter infrastructure.
Storage primitives place two distinct calls for on a CPU. First, inside every knowledge stream, encryption, integrity checking, restoration, compression, and decompression should full rapidly earlier than subsequent storage processing can proceed, making sustained per-core efficiency vital. Second, throughout the system, these operations run over many concurrent streams and repeatedly transfer knowledge by means of caches and reminiscence, making bandwidth and predictable latency equally vital.
Vera addresses each necessities. The Olympus core combines vast instruction throughput, superior department prediction, deep out-of-order execution, and vector and cryptographic sources to assist every core maintain instruction throughput throughout control-heavy and data-processing code.
NVIDIA Spatial Multithreading, the monolithic compute die, SCF, unified L3 cache, and high-bandwidth SOCAMM2 reminiscence assist hold energetic cores provided with knowledge whereas decreasing thread-to-thread interference and supporting extra predictable knowledge entry below load. Collectively, these capabilities assist clarify the measured positive factors throughout encryption, integrity checking, parity calculations, compression, and the multi-stage storage pipeline.
This permits the BlueField-4 STX Storage Processor to maintain extra CPU-side storage processing throughout concurrent knowledge streams with out proportional will increase in CPU sources, energy, and cooling.
Measuring foundational storage efficiency
Storage duties execute repeatedly throughout storage learn, write, and restoration paths. Their throughput and effectivity assist decide whether or not CPU-side processing retains tempo with SSDs and networks or turns into the limiting stage within the knowledge path. The storage primitive microbenchmarks on this submit isolate these features to measure the processor’s contribution. They present the CPU efficiency and headroom obtainable in Vera for constructing higher-throughput, extra environment friendly storage companies.
Every take a look at runs inside a single course of utilizing knowledge already held in reminiscence. The exams exclude file I/O, disk efficiency, networking, command startup, and external-device bottlenecks until in any other case recognized. The benchmark set makes use of widespread libraries, together with OpenSSL, Zstandard, and LZ4, together with comparable implementations optimized to make use of the native directions obtainable on Arm and x86 processors.
A purpose-built take a look at framework runs every workload constantly and controls buffer sizes, thread counts, CPU placement, timing, correctness validation, and consequence assortment. The algorithms and most of the software program implementations are broadly adopted. The take a look at framework applies the identical workload definitions and controls throughout Vera and x86, enabling a constant processor comparability. The outcomes might be reproduced utilizing the supply code, scripts, mounted software program variations, configurations, and consequence information, though the whole benchmark set will not be an off-the-shelf public benchmark.
These measurements set up Vera’s efficiency on the storage constructing blocks that affect safe knowledge motion, resilience, capability effectivity, and repair density. Manufacturing storage paths usually apply a number of of those operations to the identical knowledge, inflicting their CPU-processing necessities to build up.
When mixed in manufacturing storage software program, the efficiency of those particular person operations contributes to combination throughput and CPU effectivity. Larger throughput throughout particular person primitives and the multi-stage pipeline offers storage software program extra CPU headroom to maintain tempo with SSDs and networks, assist concurrent knowledge flows, and apply important knowledge companies effectively. Finish-to-end testing remains to be required to quantify full storage-system or GPU-performance outcomes.
Securing extra AI knowledge with sooner encryption
AI factories course of delicate info, together with mannequin property, enterprise data, agent context, prompts, outputs, and buyer knowledge. AES-128 is broadly used for data-at-rest and data-in-flight encryption. Encryption sits immediately within the write path, the place its throughput can decide how a lot safe knowledge a platform can course of every second earlier than encryption turns into a bottleneck. Vera delivers as much as 1.43x greater AES-128 encryption throughput than the x86 CPU used for comparability.
Decryption performs the corresponding operation on the learn path andVera delivers as much as 1.29x greater AES-128 decryption throughput than the x86 CPU used for comparability.
Larger encryption throughput allows storage programs to safe extra knowledge with out constraining writes, whereas sooner decryption reduces the time required to return protected knowledge to brokers, functions, or accelerators. These assist storage programs defend and return rising volumes of AI knowledge with out consuming an growing share of the storage-performance price range.
Defending and recovering AI knowledge sooner
Storage platforms use erasure coding to guard knowledge when drives, nodes, or knowledge fragments change into unavailable. Reed-Solomon encoding creates redundancy, whereas restoration makes use of that redundancy to reconstruct lacking or corrupted knowledge. Encoding usually happens on the write path. Restoration happens throughout rebuild, restore, or a degraded learn. Their efficiency outcomes can differ as a result of these operations use totally different compute and memory-access patterns.
Vera delivers as much as 3.26x greater Reed-Solomon throughput than the x86 CPU in a restoration workload.
Larger Reed-Solomon throughput allows storage programs to write down protected knowledge and reconstruct lacking knowledge sooner. In chosen effectivity measurements, Vera additionally completes extra safety work throughout the obtainable CPU energy envelope, serving to shorten rebuilds and scale back competition with regular knowledge companies.
Validating knowledge integrity at greater throughput
Knowledge should stay appropriate as it’s moved, saved, and retrieved. Cyclic redundancy checks, or CRCs, create checksums that storage programs use to detect unintentional corruption.
CRC and Reed-Solomon carry out complementary roles. CRC detects that knowledge now not matches the anticipated consequence. Reed-Solomon supplies the redundancy used to reconstruct lacking or corrupted info.
A storage system might calculate CRCs on each the write and skim paths, together with whereas copying knowledge between buffers. As knowledge volumes develop, these checks can eat a significant share of CPU sources.
Vera delivers as much as 3.67x greater CRC32C throughput than the x86 CPU.
Larger CRC32C throughput allows storage programs to validate extra knowledge with out integrity checking limiting reads or writes. For agentic workloads, this helps return dependable context, persistent reminiscence, and enterprise knowledge with much less CPU-side processing delay.
Agentic AI creates rising volumes of context, logs, checkpoints, retrieval knowledge, intermediate outputs, and protracted reminiscence. Compression reduces the storage capability required for this knowledge and the bandwidth wanted to maneuver it. Decompression restores the information when it’s learn. Compression might run inline or after knowledge is written, relying on the storage structure, whereas decompression is usually required when compressed knowledge is learn. Compression throughput impacts how rapidly knowledge might be diminished, and decompression throughput can have an effect on how rapidly storage programs return knowledge to agentic AI functions. The next benchmarks measure these operations independently.
Vera delivers as much as 3.29x greater compression throughput than the x86 CPU, whereas sustaining its benefit throughout the measured thread counts.
The decompression benchmark measures the corresponding operation when compressed knowledge is learn. Vera delivers as much as 1.72x greater decompression throughput than the x86 CPU below concurrency, with its benefit growing as extra employee threads run in parallel.
Compression efficiency varies by algorithm, knowledge traits, compression stage, buffer dimension, and thread depend, so these outcomes apply particularly to the measured workloads. Larger compression throughput allows storage programs to scale back the quantity of knowledge written, saved, and transferred whereas sustaining CPU-side processing efficiency.
Larger decompression throughput helps storage programs return decompressed knowledge to brokers and functions extra rapidly. Collectively, these capabilities decrease stress on storage capability and bandwidth whereas enabling every processor to compress and decompress extra knowledge as agentic AI workloads develop.
Accelerating a multi-stage storage write path
Storage programs not often execute knowledge companies independently. A safe write path might compress knowledge to scale back its footprint after which encrypt it earlier than it’s written. The benchmark set features a memory-resident pipeline that applies compression adopted by encryption to every knowledge buffer. Not like the previous benchmarks, which measure particular person operations, this take a look at measures complete pipeline throughput when two CPU-intensive storage features execute in sequence.
Vera delivers as much as 3.21x greater pipeline throughput than the x86 CPU used within the benchmark for the two-stage compression and encryption pipeline.
This consequence reveals that Vera’s efficiency benefit extends past the person operations measured earlier to a multi-stage sequence consultant of a storage write path that reduces and protects knowledge. Larger multi-stage pipeline efficiency allows the storage system to course of extra knowledge per processor, serving to maintain write throughput as agentic AI knowledge volumes develop.
Scaling agent execution and storage with Vera
Agentic AI makes CPU execution and storage processing a part of the identical AI manufacturing facility knowledge path.
The CPU runs instruments, code, retrieval, evaluation, and data-processing steps between mannequin calls. The storage system secures, protects, validates, compresses, and returns the information these steps require. Each should scale throughout the finite CPU, energy, and cooling sources obtainable throughout the AI manufacturing facility.
Vera was designed to speed up CPU-dependent work for the agentic AI period. In NVIDIA Vera Rubin, it serves because the host CPU for NVIDIA GPUs and helps agent execution. Standalone Vera delivers as much as 1.8x greater efficiency per core in agentic instruments. In BlueField-4 STX, Vera powers the CPU-side processing utilized by AI-native storage platforms.
The benchmark outcomes present Vera accelerating encryption and decryption for safe knowledge entry, Reed-Solomon restoration for sooner reconstruction of lacking or corrupted knowledge throughout storage rebuilds and repairs, CRC32C for high-throughput integrity validation, and compression and decompression to scale back capability and bandwidth calls for and speed up knowledge retrieval.
By sustaining these features individually and inside a multi-stage write path, Vera helps AI-native storage platforms course of and return safe, dependable knowledge with much less CPU-side delay. It additionally helps extra concurrent knowledge flows and better service density with out proportional development in CPU sources, energy, and cooling. Throughout chosen workloads, Vera additionally delivers greater measured efficiency per watt, enabling extra CPU-side storage processing throughout the obtainable processor energy price range.
A standard Vera CPU structure and software program toolchain throughout compute and storage supplies a constant basis for scaling each agent execution and the supporting knowledge infrastructure.
