Samsung has launched a memory semantic SSD that combines flash and DRAM and claims a 20x performance improvement compared to traditional SSDs. Modern solid-state drives with PCIe Gen4 or PCIe Gen5 interfaces deliver extremely high sequential read/write throughput and random access performance. However, this is not sufficient for applications that require a lot of performance, as they require storage devices that are explicitly tuned to their usage patterns. So, to meet the requirements of artificial intelligence (AI) and machine learning (ML) workloads, Samsung is eyeing new CXL attached SSDs.
AI and ML applications typically process large datasets, but in relatively small sized chunks of data. As such, traditional data center grade SSDs optimized for mixed workloads are typically not the best storage devices for these types of workloads.
To handle such workloads, Samsung has developed a special memory semantic SSD. It combines a huge internal DRAM cache, traditional NAND storage and a PCIe Gen5 interface with CXL technology. Applications can write data to the DRAM cache at DRAM speeds, with low latency enabled by the CXL.mem protocol. The drive’s proprietary Samsung controller (or controllers) then transfers the data to NAND memory. The result is a 20x increase in random read/write performance compared to his traditional SSD.
CXL Cache Coherent Protocol enables memory/cache coherency and significantly reduces memory access latency by various computing devices, resulting in significantly improved performance for applications using various memory expansion modules and computing accelerators I promise to let you With memory semantic SSDs, Samsung leverages high-performance DRAM (like other Type 3/CXL.mem devices) and high-capacity NAND storage memory. Memory Semantic SSD is a unique storage solution tuned to read and write small size chunks of data very fast.
Samsung has not announced which versions of the CXL protocol it supports, but the newly announced CXL 3.0 revision supports mixed-media (DRAM, storage, etc.) devices that can be attached to the fabric. However, these new Global Fabric Attached Memory (GFAM) devices are a new definition, so Samsung’s devices meet the general description of this type of device, but extend to future generations of SSDs. GFAM function may not be seen.
Samsung doesn’t provide details or specs for its Memory Semantic SSD, so we don’t know what kind of DRAM, 3D NAND, and controller they use. In fact, the company doesn’t disclose actual performance numbers for the drive.
Presumably, this is because Memory Semantic SSDs are geared for next-generation platforms that are a quarter away (such as AMD’s Genoa and Intel’s Sapphire Rapids), and Samsung is unable to disclose relevant performance numbers. Another reason Samsung doesn’t publish raw performance numbers or even specs is that the Memory Semantic SSD is also a quarter away from him. Therefore, some specifications may change before they become available.
