With the continuous improvement in processor speeds, especially in the competitive landscape of CPUs and operating systems over the past few years, CPU performance has largely followed Moore's Law. However, in the last decade, CPU speed improvements have slowed significantly. In my view, there are two main reasons for this: one is the manufacturing process, and the other is memory access latency. Data stored in memory can be difficult to load into the CPU core quickly. Additionally, as applications become more complex, memory I/O speed has become a critical bottleneck. So, has there been any breakthrough in memory I/O speed recently?
Last month, Everspin introduced their latest development: the nvNITRO NVMe storage accelerator card based on MRAM technology. What stands out is its impressive I/O speed—reaching up to 1.46 million IOPS for a random 4KB 70/30 read/write mix. This is quite surprising and seems to be among the fastest IOPS available today. For comparison, Intel’s P4800X Optane SSD card only reaches around 500,000 IOPS under similar conditions. The reason behind this high performance lies in several key factors. First, the card uses Everspin’s latest 1Gb ST-MRAM, which is compatible with DDR4 and SDRAM interfaces. Second, it includes dedicated MRAM storage control IP blocks that support NVMe 1.1+. Most importantly, it leverages the Xilinx Kintex UltraScale KU060 FPGA to implement the MRAM controller, connecting it via PCIe Gen3 x8. This setup enables ultra-fast I/O access.
The nvNITRO card is expected to be available in Q4 2017, with capacities of 1 or 2GB. It's definitely something to watch out for. Unlike traditional NVMe cards, which often rely on NAND flash or other volatile memory types, Everspin’s ST-MRAM offers non-volatility without the need for a backup power supply. Moreover, ST-MRAM has exceptional endurance, allowing unlimited write operations without degradation. This eliminates the need for wear leveling algorithms used in NAND flash, and the performance remains stable even after long-term use.
From the performance chart, it's clear that the write speed of ST-MRAM is nearly as fast as DRAM. This is one of the key reasons why the nvNITRO card achieves such high read and write speeds. Another notable feature is the ability to customize the card by writing your own RTL code onto the programmable FPGA. This means users can tailor the functionality of the card without increasing the BOM cost, making it highly flexible for specialized applications.
In summary, the focus of this article was on the impressive I/O speed of the nvNITRO card and the technical reasons behind it. When combined with previous discussions, it becomes clear that the use of FPGA plays a crucial role in achieving such performance. FPGAs offer flexibility in implementing various data interfaces and controllers, giving users more control over system design. As we look ahead, I believe more high-performance FPGA designs from companies like Xilinx will emphasize flexibility and customization even further.
600-800V Series Electric Compressor Three-phase Sealed Terminal
Shenzhen Capitol Micro-Electronics Co.,LTD , https://www.capitolgtms.com