SK hynix, one of the world’s largest semiconductor and DRAM makers, has revealed that it has working samples of its new generation memory modules designed for HPC and servers. Called Multiplexed Combined Rank (MCR) DIMMs, this technology allows high-end server DIMMs to operate at a minimum data rate of 8 Gbps. This represents an 80% increase in bandwidth compared to his existing DDR5 memory product (4.8 Gbps).
Generally, the most common way to ensure higher throughput performance with DIMMs is to increase the memory bus (and chip) clock speed. However, this strategy is not without its drawbacks, and aiming to find a more comprehensive way to do this, SK hynix is working with both Intel and Renesas to create Multiplexer Combined Rank DDR5 DIMMs. Did.
Combining Intel’s unannounced MCR technology for server chips with Renesas’ expertise in buffer technology, SK hynix’s DDR5 MCR DIMMs deliver 66% more bandwidth than traditional DDR5 DIMMs, delivering an impressive 8 Gbps/ It claims to have the bandwidth of the pin (DDR5-8000). SK hynix himself claims that MCR DIMMs will be “at least” 80% faster than his DDR5 ones out there today, but how this number is reached is not quantified. .
The technology behind MCR DIMMs is interesting. Because two ranks can be used simultaneously instead of one, you can effectively combine two sets/ranks of memory chips, doubling the effective bandwidth. Unfortunately, the details beyond this are slim and unclear. In particular, SK hynix claims that the MCR “can transfer 128 bytes of data to the CPU at a time”, but looking at the picture of his DIMM provided, it doesn’t seem like much. Pins to support physically wide memory buses.
Supposedly, SK hynix and Intel serialize memory operations for both ranks of memory within a single DDR5 channel, allowing two ranks to achieve a cumulative effective bandwidth of 8Gbps. This is supported by the use of a Renesas data buffer chip shown on his DIMM in SK hynix’s photo. Conceptually, this is not far from Load Reduced DIMMs (LRDIMMs) either. LRDIMMs use a data buffer between the CPU and memory chips, but it’s difficult to tell how far apart.
What’s more interesting is that this design puts a lot of reliance on the physical memory bus and the ability of the host controller (CPU) to run at DDR5-8000 (and higher) speeds. Usually the bottleneck in getting more memory bandwidth in server-grade systems is the memory bus in the first place. It has to run slower to accommodate more memory. So going the route that requires such a fast memory bus is an entirely different approach. In either case, being able to run his DIMMs at DDR5-8000 speeds in a server is a huge boon to memory bandwidth and throughput. This is because today’s manycore chips are often lacking.
SK Hynix partnered with Intel through MCR technology and uses Renesas’ buffer technology, which makes MCR look like an Intel proprietary technology, at least at first. As part of SK hynix’s press release, Intel said, “We look forward to bringing this technology to future Intel Xeon processors to support standardization and multi-generational development efforts across the industry.” . For the time being, this appears to be still an actively developing technology and SK hynix has not disclosed anything about availability, compatibility or pricing.
SK Hynix doesn’t say much about MCR DIMMs being twice as fast as traditional DDR5 memory, but this product is designed for the High Performance Computing (HPC) and server industries, where MCR DIMMs are It is unlikely to be seen in any industry. Create forms in consumer-based systems. We hope to learn more in the not too distant future.
sauce: SK hynix