AMD and NVIDIA today are not companies of graphics cards, processors or gaming, no, not at all, they are companies of Deep Learning and AI. The definition is literal and no matter how much revenue says otherwise, you just have to look at the efforts to implement and take advantage of the Tensor Cores or RT Cores to realize that in gaming we import what is left over from the other sectors. Proof of this is the evolution that AMD undergoes internally, where we now know of the existence of the MI300, its chiplets and all its surprises.
It was a matter of time before it arrived, it is a movement that Intel already designs, AMD already executes and NVIDIA is behind schedule. We are talking about the incorporation of high-performance CPU and GPU in the same chip, in the same package. NVIDIA can’t get into designing x86 CPUs for obvious reasons, but what about buying ARM? It is another story and how everything is focused on efficiency 128 ARM Cores powered by NVIDIA and GPUs Ada Lovelace It makes more than one heart beat just thinking about it. Well, AMD already has it done.
The AMD MI200 is on the way, its successor will break the molds of technology
There’s a server socket called “SH5” that features CPU with CPUID 0xA80F00… called “MI300”
– ExecutableFix (@ExecuFix) September 16, 2021
We tend to think that we are up to date and the truth is that this is a very short-sighted vision for how a giant of semiconductors thinks today. The projects and technologies are perfectly organized leaps from 8-10 years before and only a lithographic stop shortens these figures.
AMD has something ready that is going to revolutionize the market, something that has received the name of SH5 and which is internally called MI300, a direct successor to the MI200 that has not yet been launched. If already this GPU for Exascale is technically crazy (220 CUs and 128 GB of HBM2E), its successor is a revolutionary change in itself to a level that the chiplets of Zen architecture are child’s play.
AMD Instinct MI300, chiplets of the future with Zen 4, CDNA 3 and HBM2E
There is not much information, simple brush strokes, but they are enough to cause cardiac arrhythmias just thinking about it. The supposed SH5 socket that will integrate the new MI300 will have several core-based chiplets Zen 4 along with 4 other chiplets with GPUs CDNA 3, and as if that were not enough, it will be integrated for the latter HBM in independent interposers.
That is, we are talking about the concept of APU but with vertical stackers through interposers and TSV in several layers, massive, gigantic, symmetrical and ready for EHP (Exascale Heterogeneous Processor).
There is no hard data beyond the fact that it will integrate 4 chiplets (or eight even with less Cores) for GPU and 4 chiplets for CPU, but judging by the slide and considering that currently HBM2E can “only” get 2 GB per die in 12 dies per stack, it is more than likely that we will see a huge number of this type of VRAM: 192 GB (24GB per stack for eight stacks total).
Therefore, we are facing the most complete chip designed to date, with extreme complexity due to continuous vertical stacks and welded interposers, as well as gross power. It will nevertheless be the second MCM design from AMD and the first to incorporate CPU and GPU cores from contiguous architectures along with high-performance memory to the market. We will have to wait for data to be filtered on it and for Intel’s response with ARC HPC.