Double read and write speeds, 1.4-second transfer time for the LLM 40GB model, and 1.8x improvement in power efficiency
Amid growing demand for enterprise storage capable of processing large volumes of data with low latency and high speed due to the advancement of AI server infrastructure, Samsung Electronics has launched a new enterprise SSD (eSSD) product featuring next-generation interface standards. PCIe (Peripheral Component Interconnect Express) 6.0 is attracting attention from the industry for its ability to reduce bottlenecks in AI training and inference workloads, as it offers twice the data transfer bandwidth compared to the existing PCIe 5.0.
Samsung Electronics announced on the 8th that it has started mass production of the PCIe 6.0-based enterprise eSSD 'PM1763'.
It features 9th generation V-NAND and a new controller based on a 4nm process to improve performance and power efficiency, and is available in three capacities: 4TB, 8TB, and 16TB.
Based on the 16TB model, the sequential read speed is up to 28,400 MB (megabytes) per second and the write speed is up to 21,900 MB per second, which is about twice the level of the previous model, the 'PM1753'.
Samsung Electronics explained that this speed allows for the transmission of a 70B (4-bit quantization standard, approximately 40GB) file of the large language model (LLM) Llama-3 in about 1.4 seconds, which can increase AI processing efficiency by reducing data latency between the accelerator and the processor.
The company stated that power efficiency has improved by more than 1.8 times compared to the previous model, which can contribute to reducing data center operating costs.
The cooling design was also tailored to the next-generation AI server environment.
It was announced that by adopting a Direct-to-Chip (D2C) liquid cooling method that directly mounts a cold plate to the device, it is possible to operate for extended periods without performance degradation even under high load conditions.
In terms of security, the Post-Quantum Cryptography (PQC) encryption algorithm was applied to enable defense against quantum computing-based attacks.
In addition, through TDISP (TEE Device Interface Security Protocol) technology, in a virtualized environmentIt is equipped with a structure that blocks unauthorized external access.
Choi Jang-seok, Senior Vice President of the Product Planning Team at Samsung Electronics’ Memory Business Division, stated, “The PM1763 meets the next-generation AI platform requirements of global clients and has completed product verification,” adding, “It will become a core solution that supports the efficient operation of clients’ AI models through expanded memory capacity.”