Samsung SmartSSD CSD PM991 4TB

The Samsung SmartSSD Computational Storage Drive (CSD) PM991 4TB represents a fundamental shift in storage architecture by embedding a high-performance FPGA-based compute engine directly within the NVMe SSD enclosure. Rather than routing raw data across the PCIe bus to host CPUs for processing, the PM991 executes data transformation, filtering, compression, decompression, and AI inference kernels locally on-drive. This near-data processing paradigm collapses memory bandwidth bottlenecks and dramatically curtails the volume of data traversing the host interconnect, yielding measurable reductions in host CPU cycles consumed during analytics and inference workloads. The drive adheres to the NVMe CSD command set extensions, enabling standards-compliant integration with Computational Storage aware software stacks.

Built on Samsung's proven V-NAND flash technology and coupled with an advanced NVMe controller, the PM991 delivers enterprise-class sequential read and write throughput alongside the computational fabric necessary for in-storage processing. The integrated FPGA is programmable via industry-standard HDL and high-level synthesis toolflows, allowing enterprises and ISVs to deploy custom acceleration kernels tailored to genomics pipelines, log analytics, columnar database scans, and real-time sensor fusion. Samsung provides a CSD SDK that abstracts FPGA programming complexity, offering pre-built function libraries for common data operations such as regular expression matching, data filtering, compression (LZ4, Zstandard), and floating-point neural network inference.

Targeted at hyperscale data centers, enterprise analytics clusters, and intelligent edge inference nodes, the PM991 4TB is particularly compelling in scale-out deployments where per-server CPU headroom is constrained. By offloading pre-processing and inference tasks to the storage tier, operators can increase rack density, defer CPU platform upgrades, and achieve lower end-to-end query latency on datasets that previously required data movement to dedicated accelerator nodes. The drive ships in a standard U.2 form factor, ensuring broad compatibility with existing server infrastructure and front-loading NVMe backplanes.

Ideal for

• In-storage columnar database scan acceleration for OLAP workloads, executing predicate pushdown and projection filtering on-drive to reduce data transferred to query engines such as Apache Arrow Flight and Trino.
• Edge AI inference pipelines where pre-trained neural network models are loaded onto the FPGA fabric to perform real-time classification or anomaly detection on streaming sensor or log data without host CPU involvement.
• Genomics and bioinformatics sequence alignment preprocessing, offloading FASTQ parsing, quality trimming, and k-mer indexing directly at the storage layer to accelerate pipeline throughput.
• Security analytics and SIEM platforms leveraging on-drive regex and pattern-matching kernels to filter and classify raw log ingestion streams before escalating relevant events to the host for correlation.
• Hyperscale object storage tiering with on-drive transparent compression and decompression (LZ4/Zstandard), reducing effective write amplification and reclaiming usable NAND endurance.
• Content delivery network edge nodes performing real-time transcoding metadata extraction and content fingerprinting at the storage device level to minimize CPU load on high-density serving infrastructure.

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.