Samsung will make the next-generation Mach-1
Artificial intelligence (AI) chips for Naver Corporation, a deal worth up to 1 trillion won ($752 million USD). With its new deal with Samsung, south Korean search giant Naver will significantly reduce its reliance on NVIDIA for its AI processors. Samsung's System LSI business division has agreed to supply AI chips to Naver, with the two companies in "final talks to fine-tune the exact volume and prices," according to "people familiar with the matter," reports KED Global.
Samsung expects the price of the next-gen Mach-1 AI chip to be around 5 million won ($3756 USD or so) with Naver wanting to receive between 150,000 and 20,000 units of its new AI accelerator according to the same sources. Naver is a leading Korean online platform giant, where it will use the next-gen Mach-1 AI chips in its servers for AI inferencing, replacing the chips that it received from NVIDIA. Naver isn't Samsung's first customer, with Meta also switching things up and getting Samsung Foundry to build its next-gen AI chips, moving away from Taiwan-based TSMC (Taiwan Semiconductor Manufacturing Company). Samsung is also in discussions with
Microsoft to make new chips, too. The head of Samsung's semiconductor business, Kyung Kye-hyun, said during the company's recent annual general meeting on Wednesday that the new Mach-1 AI chip is still under development. A prototype is expected to be mass-produced by the end of the year. Naver's new Mach-1 AI accelerator is a form of a system-on-chip (SoC) that reduces the bottleneck between the GPU and HBM, according to Samsung. Kyung said that the Mach-1 is a product specified to fit the transformer module. He said: " By using several algorithms, it can reduce the bottleneck phenomenon that occurs between memory and GPU chips to one-eighth of what we are witnessing today and improve the power efficiency by eight times. It will enable large language model inference even with low-power memory instead of power-hungry HBM ". Mach-1 will combine Samsung's proprietary processors and low-power (LP) DRAM chips, unlike NVIDIA AI accelerators that feature GPUs and HBM chips. Mach-1 has fewer data bottlenecks, and consumes just one-tenth the power of NVIDIA's AI GPUs.