In the election of the RAM For the CPU, especially in high-performance processors and for high-performance computing (HPC) applications, bandwidth has become a crucial factor. With the increase in artificial intelligence (AI) workloads and new long SIMD instructions, RAM performance needs to be further improved. In this context, AMD has collaborated with JEDEC to develop a solution called MR-DIMM (Multi-Ranked buffered DIMM).
What is JEDEC?
AMD is a fairly well-known company, but… what is JEDEC? Well, JEDEC stands for Joint Electron Device Engineering CouncilJEDEC is a global industry standards organization dedicated to establishing and promoting standards in the semiconductor industry. Founded in 1958, JEDEC works in collaboration with electronic device manufacturers, suppliers, and end-users to develop technical standards in areas such as RAM, integrated circuits, chip packages, semiconductors, and other electronic components.
JEDEC is responsible for defining specifications and standards related to the manufacturing, design, testing and use of memoriesThese standards help ensure compatibility, reliability, and interoperability between different electronic devices and systems, facilitating the development and adoption of technologies in industry.
RAM as a bottleneck
La RAM has become a major bottleneckGiven the gap between the advancements in processor performance and that of RAM, a situation has arisen where RAM speed becomes a limiting factor in the overall performance of a computer system. This occurs when the CPU's processing speed is faster than the speed at which RAM can supply the necessary data.
When RAM becomes a bottleneck, there may be a decrease in system performance, as the CPU has to wait for the data needed to execute tasks. This can result in longer loading times, delays in application launch, and a reduced ability to handle memory-intensive tasks, such as graphics processing or multitasking.
To overcome this bottleneck, many different measures can be taken, but one of them is to research new memory standards that can contribute more speed and bandwidth.
DDR vs SDRAM memory
The memory DDR (Double Data Rate) and SDRAM (Synchronous Dynamic Random-Access Memory) These are two types of memory technologies used in computer systems. The move to DDR was a significant step forward, but it's still insufficient.
DDR memory is generally faster than SDRAM. This is due to its ability to transfer data on both the rising and falling edges of the clock, which doubles the effective speed compared to SDRAM, which only transfers data on the rising edge. And now MR-DIMM aims to achieve something similar to that leap, but in a different way…
Why is MR-DIMM needed?
As I mentioned earlier, it is crucial to address the problems that arise. But, What are the possible solutions?
- One option is to increase the recording speed, but this is limited due to signal integrity.
- Another alternative is to add more memory channels to the CPU, although this considerably increases the complexity of both the platform and the processor itself. Despite this, both AMD and Intel have followed this route.
- Another strategy is to use a new memory standard that reduces latency and improves data transfer capacity, such as the current move toward DDR5.
Furthermore, a more ingenious approach can be taken and a new way of accessing existing memory can be invented. This is precisely what AMD and Intel have done: AMD proposed HBDIMMWhile Intel developed MCR-DIMM. Fortunately, there will be no divisions in the industry on this aspect, as JEDEC has collaborated with AMD to develop the MR-DIMM standard based on HBDIMM.
The new MR-DIMM modules They achieve double the data transfer speeds of the same hardware in standard DDR5 mode. This is achieved by simultaneously accessing two memory ranks, either on a single module or a pair of modules, combining two 64-bit accesses into a single 128-bit data path to the CPU.
Technical characteristics of MR-DIMM memory
| MR-DIMM | Data speed |
|---|---|
| Gen1 | 8.800 MT / s |
| Gen2 | 12.800 MT / s |
| Gen3 | 17.600 Tm/s |
The first iteration of MR-DIMM modules will offer speeds of data transfer up to 8800 MT/sJEDEC expects MR-DIMMs to progressively improve, reaching speeds of 12,800 MT/s and eventually 17,600 MT/s. However, third-generation MR-DIMMs likely won't be available until after 2030, so this is a long-term project.
Although the roadmap for MR-DIMMs does not provide precise details about the arrival of The first generation, Intel Granite Rapids processors and AMD EPYC Turin (Zen 5) processors are expected to be released in 2024Although there's no official confirmation, it's plausible that Turin could also take advantage of MR-DIMMs, given AMD's recent commitment. Therefore, there's a chance MR-DIMMs could also arrive in 2024.
As for other computers, like the ones we have at home, it's unclear whether this technology will eventually be implemented. However, it would be an interesting solution for a performance boost. But for now, it only seems to be available in the supercomputing field...