Currently, Shenzhen POWEV Electronic Technology Co., Ltd. (hereinafter referred to as "POWEV") is developing a new - generation memory, **MRDIMM**, which is more suitable for professional AI applications. Through the upgrade of memory capacity and performance, MRDIMM will greatly enhance the scale and efficiency of AI computing, and further reduce the hardware costs of small and medium - sized enterprises in AI R & D, operation and maintenance, and application.
In the past, servers and workstations mainly used RDIMM memory. However, with the rapid improvement in the performance of computer processors and graphics cards and the development of AI computing capabilities, traditional RDIMM memory has struggled to meet the requirements of professional AI application fields.
As a new type of memory chip, High Bandwidth Memory (HBM) has now evolved to HBM5. It has a relatively high cost and requires advanced hardware solution development capabilities. Generally, it is used in smaller and more large - scale AI application terminals.
Powev is currently dedicated to researching MRDIMM memory. This product can be widely used in servers and workstations, significantly improving the operating efficiency of AI and reducing the development, operation, maintenance, and application costs of AI for small and medium - sized enterprises.
The core components of MRDIMM memory include the Multiplexed Registered Clock Driver (MRCD) and the Multiplexed Data Buffer (MDB). It doubles the data throughput through parallel transmission. By operating two memory ranks simultaneously and combining multiplexing technology, the data transfer rate is increased to twice that of the standard DDR5 RDIMM.
The advantages of MRDIMM memory are manifested in multiple aspects. MRDIMM is fully compatible with the physical interfaces and form factors of existing DDR5 RDIMMs, allowing for direct upgrades without modifying the server motherboard. Generally, as the memory capacity increases, it becomes more difficult to improve performance. MRDIMM's dual-rank operation doubles the bandwidth, making it a technology that enables the coexistence of large capacity and high performance. The computer processor, graphics card, and memory are the three important factors affecting AI computing. With the improvement of the performance of processors and graphics cards, memory has become the bottleneck in AI computing. MRDIMM significantly alleviates this contradiction by increasing bandwidth and efficiency.
For example, in the local deployment applications of DeepSeek, ordinary users can choose a working mode with a model size below 70B, which has relatively low requirements for memory capacity and bandwidth. However, small and medium - sized enterprises that opt for a working mode with a model size above 70B will surely find it difficult to rely solely on the video memory of graphics cards. At this time, the capacity and bandwidth of the memory become crucial factors affecting the operation efficiency. Choosing MRDIMM memory will reduce the computer configuration cost for the local deployment of DeepSeek applications. Moreover, when the configuration remains unchanged and the memory is upgraded to MRDIMM, the energy efficiency of AI computing will also be significantly improved.
The MRDIMM memory being developed by POWEV has a single - stick capacity ranging from 32GB to 256GB. The first - generation supports a speed of 8,800 MT/s, the second - generation will support a speed of 12,800 MT/s, and the third - generation is expected to exceed 17,600 MT/s. According to preliminary tests, when using the first - generation MRDIMM with the Intel Xeon 6 processor, the performance can be improved by up to 33%. In AI inference tasks, the token throughput is increased by 31% and the latency is reduced by 24%. The combination of the large - capacity, high - bandwidth, and low - latency characteristics of MRDIMM with the requirements of AI computing power has significantly reduced the hardware cost of AI. In the future, it will gradually become the standard configuration for AI servers and high - performance computing.