Mach-1, an AI chip with more power and improved communication between memory and GPU | Samsung | mobiles | smartphone | TECHNOLOGY

Samsung is working on the development of its new chip Artificial intelligence (IA) next generation Mach-1, with which it will offer more power by reducing the “bottleneck” phenomenon that occurs between the processor memory and the graphics processing unit (GPU) to one eighth, of the same way that energy efficiency spending will decrease.

Currently, AI chips are playing an essential role in advancing innovation and issues related to, for example, large language models (LLM). Therefore, in a market led by processor development companies such as Nvidia, Samsung continues to advance to offer new technologies that are up to the task.

In this context, Samsung has announced that it is working on its first next-generation AI processor, which it has called Mach-1, and that it plans to launch early next year.

This was detailed by the director of Samsung Semiconductors, Kyung Kye-hyun, within the framework of Samsung’s annual general meeting held in Seoul (South Korea), as The Korea Economic Daily has learned.

Specifically, Kye-hyun has pointed out that current AI systems suffer from performance and power issues due to “memory bottlenecks.” In order to find a solution, he has indicated that Samsung is establishing an Artificial General Intelligence (AGI) computing laboratory, that is, a type of automatic system that has the ability to successfully carry out any human intellectual task, such as reasoning. , learn or even plan for the future.

In this way, Samsung intends to use the AGI computing laboratory, as well as continue innovating in AI architecture, to introduce these capabilities to the Mach-1 chip.

Likewise, the Mach-1 chip will function as an AI accelerator that will improve the transmission of information between the GPU and Samsung’s high-bandwidth memory (HBM) microprocessors, responsible for achieving faster processing speed to manage AI tasks.

To improve this transmission, the new chip will include “several algorithms” that will reduce the bottleneck phenomenon to “one eighth”, compared to current AI processors. Likewise, energy efficiency will be improved up to “eight times,” according to Kye-hyun.

In this way, Mach-1 will have the ability to allow the inference of large language models (LLM) “even with low-power memory instead of HBM, which consumes a lot of energy,” the manager stressed.

With all this, Kyung Kye-hyun has assured that the Mach-1 AI chip is currently in development and that, therefore, the company plans to have a prototype by the end of this year 2024. After this, Samsung has proposed to implement the Mach-1 chips in AI systems “early next year.”

As the same media has pointed out, this new chip is also part of Samsung’s attempt to surpass its rival SK Hynix, which recently began mass production of its high-bandwidth HBM3E1 chip. However, last February Samsung announced that it was developing the industry’s first 36GB HBM3E 12 H DRAM memory, which it claims will increase performance and capacity by more than 50 percent in this type of chip.

With all this, according to Kye-hyun, Samsung also plans to regain the number one position in chip manufacturing in a period between “the next two to three years.”