d-Matrix Corsair: 256GB of LPDDR for AI Models
Hello you fine Internet folks, Today is the last of our Supercomputing 2024 coverage, but last does not mean least. Today we are going to be covering a new AI startup company called d-Matrix. d-Matrix first product is called the Corsair and it uses the Microscaling formats from the Open Compute Project with the Microscaling formats being adopted by Nvidia in the Blackwell series of GPUs as well as in Microsoft’s MAIA 100 AI chip.
Looking at the chip itself, it’s very remeniscent of AMD’s Naples series of CPUs with 4 identical rectangular dies on a package. However that is where the similarities between Naples and Corsair end.
Where Naples is a general purpose CPU meant to do all sorts of computing, Corsair is strictly for Machine Learning and Low-Precision Computing.
The basic building block of the d-Matrix architecture is the Digital In-Memory Compute (DIMC) core. The DIMC core integrates a multiplier into a memory bit cell which reduces latency and the energy cost of access memory since you don’t need to move the data anywhere because your multiplier in already in your memory.
