Computer

This chip for supercomputers shows the future of PCs

The high manufacturing costs and certain limits that did not exist before make it necessary to look for alternative solutions when designing new processors when building complex systems. One of the trends that sounded the most on paper is the transfer of the internal intercommunication of a chip to what we call an interposer in a 2.5D integrated circuit, that is, one made up of several small chips instead of a single large one. The MI300 is the first product to deliver on that promise, and although it’s a server processor and therefore high-performance computing, it’s interesting to take a look at it for the future.

AMD achieves fully coherent memory in the MI300

One of the particularities that video game consoles have is the fact that their memory space is completely unified on a physical level, that is, all the elements share access to the same RAM memory chips. However, what is done is to separate the storage space into two different halves to avoid conflicts regarding memory coherence. Ideally, however, all the space should be shared, as this would prevent the use of copying processes from one part of the RAM to another.

CDNA MI300

That is, even having physically shared memory, it is necessary to use the DMA unit to copy from CPU space to GPU space and vice versa. The unification in addressing of both memory pools is one of the things that Lisa Su’s company has been talking about, and it’s not a joke, for a decade. However, they sell it as one of the upgrades to look out for on the MI300. In any case, one of the peculiarities of the CXL standard is to grant this capacity to PCI Express, but this has a problem that we will discuss later.

Why is this a clue to the future?

Until now we have seen the implementation of a fully coherent memory system with a monolithic chip where the CPU and GPU share the same memory controller. If you look at it, the MI300 is made up of several chips with its own memory controller and yet there is a complete coherence mechanism. It is not because of the use of HBM memory, but because of the changes made by AMD in the memory controller, especially of the GPU of this mastodon for servers.

MI300 chip

Of course, the grace lies in the fact that the integrated memory controller is located and it is not any of the different CPU and GPU chiplets, but rather in the interposer, and that is that all of them are mounted on another much larger chip. , in charge of communicating them with each other and with the HBM memory. Now, if we think about it, since said interposer is a separate piece, then it can be easily transferred to another type of composition. How about a future video game console? (https://eluminoustechnologies.com/)

The trend with rising costs is to break up chips into several different ones, it is not an AMD whim and it has been shown that certain components do not scale as well using smaller transistors, hence separating these functions into separate chips made with cheaper nodes. In addition to the above, the MI300 Interposer also includes the so-called Infinity Cache as LLC, which is key to achieving a fully coherent system.

Will we see it on PC?

A UMA configuration on a computer would not be impossible, however, this would mean selling processor, graphics and shared RAM in the same package and it does not seem to be economically cheap. GDDR6 memory is used on the console due to the cost structure, but on PC the use of said memory would be deadly, due to having a much higher latency than conventional RAM. That is, using the same processor performance is lost. The DDR5? Even faster than DDR4, it doesn’t have enough bandwidth to be viable for a high-caliber integrated GPU.

Future Render Processors

So in the end all they are left with is HBM memory, extremely expensive to put into a home system. So guessing we only have game consoles and a future Radeon graphics card, maybe RDNA 4? In which the function of the small chiplets in the RX 7000 size is transferred to the interposer, giving the possibility of breaking down the main chip. In short, what AMD has done is move the classic Northbridge, what they call IOD in their marketing jargon, to the interposer.

Related Articles