'1st Place Reclaim' Declared by Samsung Electronics: Developing "Maha1"... Attempt to Overturn AI Semiconductor Market

Regular General Meeting Management Strategy
Year-end Yangsan Plan Disclosure
"Memory Bottleneck Solving AI Inference Chip"

Samsung Electronics announced that it plans to develop its first artificial intelligence (AI) semiconductor supporting large language models (LLM), called 'MACH-1,' by the end of this year and launch it in early next year. The company aims to increase customer orders through the MACH-1, a general-purpose artificial intelligence (AGI) semiconductor solution. The plan is to seize the lead in the AI semiconductor market from SK Hynix by using MACH-1, an AI accelerator (a semiconductor package specialized for AI training and inference) that does not require high-bandwidth memory (HBM). Samsung Electronics' goal is to regain the world's number one position in semiconductors within 2 to 3 years.

Kyung Kye-hyun, President and CEO of Samsung Electronics' Device Solutions (DS) Division, revealed on the 20th during the 55th Annual General Meeting of Shareholders business strategy presentation held at the Suwon Convention Center in Gyeonggi Province that they are developing the MACH-1 chip for LLMs. This is the first time President Kyung has mentioned MACH-1. During the 'Dialogue with Shareholders' following the regular shareholders' meeting, he said, "We have completed technology verification of MACH-1 through programmable semiconductors (FPGA), and we are currently designing the system on chip (SoC). We expect to produce the chip by the end of the year and see a system composed of our chip in early next year."

AGI possesses intelligence close to or surpassing human intelligence. It requires a large amount of data and suffers from severe memory bottlenecks. Within AI accelerators, graphics processing units (GPUs) and HBM exchange enormous volumes of data to advance AI capabilities. This causes bottlenecks, which refer to the phenomenon where memory cannot properly handle large amounts of data, resulting in slower speeds and increased power consumption.

Kyung Kye-hyun, President and CEO of Samsung Electronics DS Division. / Hwaseong = Photo by Kim Hyun-min kimhyun81@

President Kyung presented a vision to enable AI inference using low-power (LP) memory semiconductors instead of HBM through the MACH-1 system. MACH-1 reduces data bottlenecks to one-eighth and improves power efficiency by eight times. The plan is to ensure that attaching LP DRAM instead of HBM to the AI accelerator will not cause issues. President Kyung explained, "We are preparing to enable LLM inference even with low-power (LP) memory."

He added, "Existing AI systems suffer from performance degradation and power issues due to memory bottlenecks. To address this, we have established an AGI Computing Lab and will pursue fundamental innovations in AI architecture."

Meanwhile, news that Samsung Electronics is increasingly likely to supply HBM to Nvidia led Samsung Electronics' stock price to close at 76,900 won, up 5.63% from the previous day. Jensen Huang, CEO of Nvidia, said on the 19th (local time) in response to a question about "whether you are using Samsung HBM," "We are currently testing it and have high expectations."