Like Mi200, Mi300 is a family of graphic processors for data processing centers. Therefore, it cannot be said that the Mi300 has one set of specifications with the MI200. On the contrary, the Mi300 will be available in many formats, including two to eight HBM3 memory glasses. In addition, the Mi300 will have three-dimensional stacking with input-output (or basic) cells directly under new computing crystals.
It should be noted here that computing crystals for the Mi300 will allegedly will not have one type of IP block, but rather give customers several options for a choice. Under each 5-nm computing crystal is the input-output crystal. This basic tile, made using a 6-nm process, is attached to two HBM3 memory stacks. This will be the first time that AMD uses 3D matrixes located one over the other.
In terms of power, each computing crystal can consume about 150 watts. Thus, the maximum configuration can consume more than 600 W, which is not surprising, given that the Mi250X already has a specification of TDP 560 W for the OAM form factor.
Each computing matrix should have an area of about 110 mm², and an intermediate element for the largest option is 2750 mm². These computing crystals will also have about 20,000 compounds with each other, which is twice as much as Apple offers M1 Ultra for its silicon computing chips.
The release date of the AMD Instinct Mi300 is currently unknown, but there is good reason to believe that by the time the Mi300 exit on the market, the NVIDIA Hopper and Intel Ponte Vecchio should already be available.