IN Brief:
- PM1763 is available in 4TB, 8TB, and 16TB capacities with PCIe 6.0 connectivity.
- The 16TB drive reaches sequential reads of 28,400MB/s and writes of 21,900MB/s.
- Storage power, cooling, and security are becoming integral parts of AI-server architecture.
Samsung Electronics has begun mass production of the PM1763, a PCIe 6.0 enterprise solid-state drive developed for AI and high-performance computing servers.
The drive combines Samsung’s ninth-generation V-NAND with a controller manufactured on a 4nm process. Capacity options comprise 4TB, 8TB, and 16TB, allowing system builders to balance storage density, bandwidth, endurance, and the number of drives connected to each processor or accelerator platform.
The 16TB model provides sequential read performance of up to 28,400MB/s and sequential writes of up to 21,900MB/s, more than twice the performance of the preceding PM1753 generation. Samsung calculates that a 40GB large-language model can be transferred in approximately 1.4 seconds under suitable system conditions.
Power efficiency is more than 1.8 times higher than the previous drive, while the mechanical and thermal design has been optimised for direct-to-chip liquid-cooled servers. Storage still requires suitable conduction paths or residual airflow, but compatibility with liquid-cooled architectures reduces the risk of leaving SSDs as isolated air-cooled components inside otherwise liquid-managed racks.
Security functions include support for post-quantum cryptographic algorithms and the TEE Device Interface Security Protocol. TDISP protects communication paths between devices and trusted execution environments in virtualised systems, where accelerators and storage may be assigned dynamically among several workloads.
AI storage moves beyond raw capacity
AI infrastructure is often discussed in terms of accelerator performance and high-bandwidth memory, yet storage determines how quickly models and datasets reach those processors. Faster GPUs deliver little benefit while waiting for information to cross a slower network or storage tier, particularly during checkpointing, model loading, retrieval, and large-scale inference.
PCIe 6.0 doubles the raw transfer rate available from PCIe 5.0 while introducing signalling and error-control changes needed to maintain reliability at higher speeds. The interface places tighter demands on signal integrity, retimers, connectors, board materials, switch architecture, and validation, and the drive’s headline throughput can only be used when the host provides enough lanes without creating contention elsewhere.
As power density becomes equally influential, several high-performance SSDs installed beside accelerators and network interfaces can add a substantial thermal load to the server. Improved efficiency reduces heat per transferred byte even as total rack consumption rises through greater drive count and utilisation.
Liquid cooling is consequently moving from an accelerator-specific addition towards a rack-level design decision. Cold plates, manifolds, leak detection, service procedures, and coolant distribution must be considered alongside memory, storage, power supplies, and network hardware. Components that cannot tolerate reduced airflow can constrain the entire mechanical architecture.
The concentration of investment around AI infrastructure, examined in the latest quarterly electronics market analysis, is spreading beyond processors into controllers, NAND, security, interconnect, power conversion, and cooling. PM1763 represents that wider system demand, with each supporting component expected to operate at higher duty cycles and exchange data at greater speed.
Because enterprise data can retain value for decades, post-quantum protection is moving into storage hardware before cryptographically capable quantum systems become available. Information encrypted today may remain sensitive after future computing advances weaken current algorithms, making device-level and platform-level security a requirement for infrastructure expected to remain in service for several technology generations.
Mass production follows validation of PM1763 for next-generation AI platforms, giving server manufacturers a production PCIe 6.0 storage option rather than an engineering sample. Effective use will still require a compatible processor, switch, board, firmware, cooling, and management stack, rather than substitution into an unchanged PCIe 5.0 system.
As storage moves closer to accelerator-class bandwidth, system bottlenecks will shift rather than disappear. Network fabric, software scheduling, metadata handling, data placement, and application behaviour must all keep pace, while the thermal and security burden expands alongside throughput.
PM1763 brings the physical drive into that transition with higher interface speed, denser flash, improved efficiency, and hardware security. Its performance will be measured less by an isolated benchmark than by the amount of useful data the complete server can place in front of its accelerators without exceeding power and cooling limits.
