The “technology verification has been completed and the design of the SoC is currently underway. Production is planned for the end of the year and presentation early next year.”
With these words, Kye Hyun Kyung, the president in charge of the semiconductor (DS) division of Samsung Electronics, announced during the 55th ordinary shareholders’ meeting held at the Suwon Convention Center, the progress made by the Korean company in the development of a proprietary AI accelerator chip, the “Mach-1”
Mach-1 has been specifically designed to meet the needs of artificial intelligence systems, inference in particular, that critical phase in which the knowledge acquired by models during training must be applied to real data to produce concrete results. Currently, the problem with AI inference is that there is a bottleneck between data transfer between GPU, CPU, and memory.
Mach-1 is expected to have a structure capable of reducing the bottleneck of data transfer speed to 1/8 of current levels. In addition, this chip will also be able to use LPDDR memory and not depend only on High Bandwidth Memory (HBM).
With the creation of this chip, moreover, Samsung obviously also intends to at least try to counter the current dominance of the AI semiconductor market by Nvidia, which just in the past few hours announced Blackwell B200, the most powerful AI chip in the world thanks to a speed of 20 petaflops of FP4 power.
Samsung Electronics is also looking to provide Nvidia with the HBM memory used in AI GPUs, but it will face stiff competition from SK Hynix. Samsung recently set up an artificial general intelligence (AGI) lab in Silicon Valley, US, to focus on the development of semiconductors for AI.