Storage plays an active role in agentic AI workflows: agents retrieve enterprise knowledge, access persistent memory, reuse key-value caches, invoke tools and generate results. Each agent step can trigger multiple storage operations, and these operations can repeat across thousands of concurrent agents. As context windows grow, storage systems must do more than simple reads and writes—they must also encrypt, compress, check integrity, and recover data without degrading application responsiveness or token generation rates.
Inference typically runs on GPUs, but agent execution, tool calls, data-management tasks and the storage services that support them run on CPUs. A single delayed storage operation can throttle the entire data flow: faster SSDs and networks cannot realize their full potential if the CPU that prepares and protects data cannot keep pace.
How Vera and BlueField-4 STX address the gap
The NVIDIA Vera BlueField-4 STX Storage Processor, part of the NVIDIA STX foundation for AI-native data platforms, brings Vera CPU performance directly into the storage data path. The same Vera CPU architecture intended to keep NVIDIA Rubin GPUs fed is used to accelerate CPU-side storage processing.
NVIDIA’s microbenchmarks report that Vera outperforms a comparative x86 CPU across storage primitives—encryption/decryption, recovery, integrity checking, compression/decompression and a multi-stage pipeline—allowing storage platforms to process more data with less CPU and power overhead and, in some cases, reducing capacity and bandwidth demands through higher compression throughput.
Vera CPU architecture: designed for storage’s dual needs
Vera includes 88 NVIDIA-designed Olympus CPU cores compatible with the Armv9.2 instruction set and supports 176 NVIDIA Spatial Multithreading threads. The design pairs these cores with the NVIDIA Scalable Coherency Fabric (SCF), field-replaceable SOCAMM2 LPDDR5X memory and a 164 MB unified L3 cache. The SCF provides up to 3.4 TB/s of bisection bandwidth; the memory subsystem can offer up to 1.2 TB/s aggregate bandwidth (up to 14 GB/s per core). These features aim to sustain strong single-thread performance and high-throughput CPU execution under bandwidth-intensive and highly concurrent workloads.
Storage primitives impose two distinct demands: within a stream, CPU work must finish quickly (favoring per-core sustained performance); across the system, many concurrent streams repeatedly move data through caches and memory (demanding bandwidth and predictable latency). Vera addresses both with wide instruction throughput, advanced branch prediction, deep out-of-order execution, vector and cryptographic resources per Olympus core, plus the SCF, unified cache and high-bandwidth SOCAMM2 memory to keep active cores fed and reduce thread-to-thread interference.
Benchmarking foundational storage primitives
NVIDIA ran memory-resident microbenchmarks that exclude disk I/O, networking and external-device bottlenecks to isolate processor contribution. Tests used common libraries (OpenSSL, Zstandard, LZ4) and optimized implementations for Arm and x86, with a consistent test framework controlling buffers, threads, CPU placement, timing and correctness validation.
Reported peak comparative results (Vera vs. the x86 CPU used by NVIDIA) include:
- AES-128 encryption: up to 1.43× higher throughput.
- AES-128 decryption: up to 1.29× higher throughput.
- Reed–Solomon recovery workload: up to 3.26× higher throughput.
- CRC32C integrity checking: up to 3.67× higher throughput.
- Compression: up to 3.29× higher throughput, maintaining advantage across thread counts.
- Decompression: up to 1.72× higher throughput under concurrency.
- Two-stage pipeline (compression then encryption): up to 3.21× higher throughput.
These measurements indicate Vera’s advantages not only on individual primitives but also on a multi-stage write path representative of secure storage flows.
Practical implications for AI workloads
- Higher encryption and decryption throughput lets storage systems secure and return more data without making the write or read path the bottleneck.
- Faster Reed–Solomon throughput can shorten rebuild and repair times and reduce contention with normal services for CPU resources.
- Increased CRC32C throughput enables validating more data without integrity checks limiting reads or writes.
- Improved compression and decompression throughput lowers storage capacity and bandwidth demands and speeds delivery of decompressed data back to agents and applications.
Overall, Vera helps AI-native storage platforms process and return secure, reliable data with less CPU-side delay, enabling more concurrent data flows and higher service density without proportional increases in CPU count, power and cooling. NVIDIA also reports higher measured performance-per-watt on selected workloads, allowing more CPU-side storage processing within a given power budget.
Conclusion and further reading
Vera was designed to accelerate CPU-dependent work for the agentic AI era: as the host CPU in NVIDIA Vera Rubin it supports GPU workloads and agent execution, and in BlueField-4 STX it accelerates CPU-side storage processing for AI-native platforms. The benchmarks show Vera improving throughput for encryption/decryption, Reed–Solomon recovery, CRC32C validation, and compression/decompression—both individually and in a multi-stage write pipeline—helping storage systems keep pace with SSDs and networks as AI data volumes grow.
For more technical detail, NVIDIA provides documents on the Vera CPU architecture and a BlueField-4 STX Storage Processor datasheet.
