AMD Aldebaran, new clues on the MCM project: two dies and CDNA 2 architecture?
For the world of GPU, first and foremost that of accelerators for data centers, the era of projects will soon open MCM (Multi-Chip-Module), where two or more GPUs on the same package will work in concert to ensure superior performance to so-called monolithic designs.
This change of philosophy will allow manufacturers to have greater flexibility in product design, with beneficial effects on the performance and consumption front: something similar, in short, to what has been seen in the CPU field with the last two generations of Ryzen.
Intel with Xe HPC, Nvidia with a project renamed Hopper (according to rumors) and AMD with CDNA 2 have all been working on these solutions for some time and the first concrete results should already be seen between the end of the year and the beginning of next year. In the case of AMD, it has long been rumored the arrival of an accelerator called Instinct MI200, codenamed Aldebaran, equipped with two GPUs on the same package.
In the past few hours, some references in a patch for the Linux kernel have revived the indiscretion, substantially confirming it. In fact, we speak of Aldebaran as a project with two dies, a die0 and a die1, where the primary die is the one responsible for managing power-related data, that is, it will regulate consumption and power limits for the entire package or at least for the two GPUs.
At the moment, the specifications of Alderaban are not known and the few hypotheses are contradictory: there are those who expect a great leap forward in performance and those who expect a more modest improvement, dictated by the novelty of the technology.
Of CDNA 2 we know that it is an “exascale supercomputer proof” project, which sees the use of the third generation Infinity Fabric interconnect as a central element. There is also talk of a more advanced production process than the 7 nanometers for the realization of the GPU, but currently it is not possible to say with certainty whether AMD refers to 5 nanometers.
In any case, the MCM design will probably also be applied to the world of desktop video cards – at least as far as AMD is concerned – starting from the solutions based on RDNA 3 architecture planned for the last part of 2022: in this regard we are talking about at least two models based on an MCM project.