Samsung Mass-Produces PCIe 6.0 Enterprise SSD Targeting NVIDIA's Next-Gen AI Platform

Samsung Electronics announced on July 8 the commencement of mass production of the PM1763, the company's first enterprise solid-state drive (eSSD) based on PCIe 6.0. This product is specifically designed as a deeply customized memory hardware solution for NVIDIA's (NVDA-US) next-generation Vera Rubin AI computing platform.

As the volume of data required for AI training and inference continues to grow rapidly, enterprise SSDs (eSSDs) capable of fast and reliable data delivery are becoming increasingly critical for AI infrastructure.

With high-speed data transfer and an optimized controller architecture, the PM1763 is expected to become a key memory solution for high-performance AI platforms. The PM1763 adopts Samsung's ninth-generation V-NAND and a newly developed 4-nanometer controller, significantly enhancing the product's performance and energy efficiency.

The PM1763 is Samsung's first eSSD to feature ninth-generation V-NAND flash memory and a 4-nanometer process PCIe 6.0 controller. The new controller has undergone architectural optimization, greatly improving data processing efficiency. The product supports the NVMe 2.1 technical standard and offers three capacity options: 4TB, 8TB, and 16TB.

Compared to the previous-generation PM1753, the PM1763 has doubled sequential read and write speeds. For the 16TB version, sequential read speed reaches 28,400MB per second, and sequential write speed reaches 21,900MB per second. At this speed, transferring a 40GB large language model takes only 1.4 seconds (estimated using Meta's Llama-3 70B parameter model).

During AI training and inference processes, SSDs are responsible for transmitting large volumes of information, including AI model data and key-value (KV) cache data, to CPUs and GPUs. By minimizing data transfer latency, the PM1763 effectively ensures that AI computing accelerators and processors maintain high operational efficiency.

The PM1763 has been specifically optimized for liquid-cooled environments in next-generation AI servers, adopting Direct-to-Chip (D2C) cooling technology. Cooling plates are directly connected to components and parts, ensuring stable performance during prolonged high-load operations without performance degradation. Additionally, its energy efficiency is 1.8 times higher than the previous generation, significantly reducing data center operating costs.

In terms of security, the PM1763 incorporates post-quantum cryptography (PQC) algorithms, providing fundamental protection against attacks from quantum computers. Furthermore, the product supports TDISP (TEE Device Interface Security Protocol), protecting data transmission channels in virtualized environments and preventing unauthorized external access and interference.