Rambus Spins 6.4 GT/s DDR5 Registering Clock Driver for Speedier Servers
The latest chip from Rambus raises the speed limit for DDR5, giving memory-hungry applications, such as data center servers, more room to run.
Marking a leap forward in DDR5 memory performance, Rambus announced the availability of its latest registering clock driver (RCD) chip enabling transfer speeds up to 6400 MT/s. As CPU frequencies and core counts increase, memory can quickly create a system bottleneck if its performance is not commensurate with the performance of the CPU, a problem that Rambus hopes to address with its RCD.
The Rambus Gen3 RCD provides a 6400 MT/s transfer speed, illustrated by the RCD eye diagram. Image used courtesy of Rambus
Especially in the case of data centers, memory bandwidth is of paramount importance in modern computing. DDR5 memory is poised to offer an incremental improvement over DDR4 in the near future while raising the theoretical maximum performance for memory in the long term. The RCD plays a crucial role in server-grade memory performance, making Rambus’s contribution an important milestone in the DDR5 journey.
In order to gain more insight into the specific benefits offered by their latest RCD, we interviewed John Eble, VP of Engineering, Datacenter Products at Rambus to get his insights on the significance of this technology for system designs leveraging DDR5 DRAM.
We’re Gonna Need a Bigger DIMM
With an ever-increasing interest in AI/ML and HPC applications, data centers are now hungrier than ever for high-performing memory. To put double data rate (DDR) DRAM into context, the DDR4 standard for computer memory is nearing its maximum theoretical transfer speed. And, as such, DDR5 is needed to allow continued scalability of memory.
The DDR5 DIMM block diagram highlights the importance of the RCD in ensuring memory performance without requiring direct intervention from the CPU. Image used courtesy of Rambus
This becomes especially true considering trends in the CPU market are pointing toward a higher number of cores in lieu of pure clock frequency, as the latter comes with decreased power efficiency. For HPC applications in particular, high-performing RAM is necessary to ensure maximum computational efficiency—a dynamic that John Eble understands well.
“As you increase the number of cores, you not only need more bandwidth for the entire chip, but you also need more capacity for the entire chip because each of those cores is wanting. There’s either a flat or slow increase in the amount of bandwidth and capacity you need per core, but then you have the number of cores increasing.”
In server-grade applications, consistent memory performance is required across all operating conditions, which is where the RCD can provide distinct advantages. The RCD acts as an intermediary between the CPU and the memory module and helps to ensure that the memory speed is maintained across temperature and computational load.
Continued Innovation in DDR5 Speed
Rambus’s latest RCD is actually the third generation of their RCD chips, offering a 33% performance improvement over its Gen1 DDR5 devices. In addition to the blazing 6.4 GT/s memory speed, the RCD uses only 1.1 V to help improve power efficiency and thermal performance in data center applications. More information is available in the DDR5 RCD 6400 MT/s product brief.
DDR5 memory offers numerous advantages compared to DDR4 including power efficiency, capacity, and bandwidth improvements. Image used courtesy of Rambus. (Click image to enlarge)
The 6400 MT/s RCD can be used in conjunction with other DDR5 chips from Rambus including its Serial Presence Detect (SPD) Hub and Temperature Sensor to allow the possibility of “smarter” memory. When discussing a DDR5 DIMM including Rambus hardware, Eble says “It has the hub and the thermal sensor, all have accurate temperature sensors, so the baseboard management controller (BMC) or the CPU can get a read of the temperature and make decisions in terms of fan speed or throttling or increasing refresh rate.”
Better, Faster, and More Efficient Memory
When browsing the latest advancements in HPC, it is easy to get caught up in the computational power of processors and neglect the effects of memory. However, memory performance can quickly strip the most powerful processor of its speed and can ultimately make system performance subpar. It’s for this reason that Eble believes that DDR5 memory will be integral in data centers across the world.
Using the analogy of the car industry, memory and CPU designs have to be designed closely aligned with one another—with planning and work done years ahead, says Eble.
“It's like we’re making components that are kind of central to, let’s say, the engine [of a car]. Memory is tightly coupled to the CPU. If you were doing a PCIe card, a CXL card, or an SSD, then there’s a little more independent work that can happen, because you can do a plug-in later on—it’s almost like an accessory. But for that CPU to really work, it needs to work closely with the memory.”
As is typically the case, the overall performance of a memory module is dependent on a variety of factors, so the inclusion of the Rambus RCD is not a guarantee of 6.4 GT/s performance. However, the continued innovation provided by Rambus appears to be enabling DIMM manufacturers to steadily increase the performance of their products.
Considering that little over 1 year has passed since Rambus released its 5600 MT/s RCD, it is natural to wonder when we may reach the 8400 MT/s limit of DDR5 memory. Further innovation is expected to enable the industry to hurdle even that milestone, if history is any indicator of the computing industry’s ability to keep evolving.