Samsung announced its Memory-Semantic SSD that combines flash and DRAM to provide a claimed 20X performance improvement compared to a traditional SSD. Modern solid-state drives with a PCIe Gen4 or PCIe Gen5 interface enable extremely high sequential read/write throughput and random access performance. But for many performance-hungry applications, this is not enough as they need storage devices explicitly tailored for their usage patterns. So, to meet the requirements of artificial intelligence (AI) and machine learning (ML) workloads, Samsung is turning to its new CXL-connected SSD.
AI and ML applications usually process large datasets, but they do it in relatively small-sized data chunks, which is why traditional data center grade SSDs optimized for mixed workloads are usually not the best storage devices for those types of workloads.
To address such workloads, Samsung developed its special Memory-Semantic SSD that combines a huge built-in DRAM cache, traditional NAND storage, and a PCIe Gen5 interface with the CXL technology on top. Applications can write data to the DRAM cache at DRAM speeds and with low latency enabled by the CXL.mem protocol. Then, Samsung's proprietary controller (or controllers) of the drive transfers data to NAND memory. The result is a 20 times random read/write performance uplift compared to traditional SSDs.
The CXL cache-coherent protocol promises to enable substantial performance uplift for applications that use various memory expansion modules and compute accelerators by enabling memory/cache coherency and dramatically reducing memory access latency by various compute devices. With its Memory-Semantic SSD, Samsung takes advantage of high-performance DRAM (like other Type 3/CXL.mem devices) and high-capacity NAND storage memory. To a large degree, Memory-Semantic SSD is a unique storage solution tailored to read and write small-sized data chunks at very high speeds.
Samsung hasn't announced which version of the CXL protocol it supports, but the newly-announced CXL 3.0 revision will support mixed-media (like DRAM and storage) devices that can be connected to the fabric. However, these new Global Fabric Attached Memory (GFAM) devices are newly defined, so although Samsung's device meets the general description of this type of device, we likely won't see the expanded GFAM functionality until future generations of the SSD.
Samsung has not provided any details or specifications of its Memory-Semantic SSDs, so we do not know what kind of DRAM, 3D NAND, and controller the company uses. In fact, the company even did not disclose the actual performance numbers for the drive.
Perhaps, that's because Memory-Semantic SSD is tailored for next-generation platforms (such as AMD's Genoa and Intel's Sapphire Rapids) that are quarters away, and Samsung cannot disclose relevant performance numbers as they are confidential for now. Another reason why Samsung does not publish even raw performance numbers or specifications is that its Memory-Semantic SSD is also quarters away. As a result, some of its specs may change by the time it becomes available.
