Every time you send a prompt via ChatGPT, data undergoes a real relay race. Information leaves memory, passes through the CPU for initial processing, then moves to the GPU for complex calculations, and travels back again. This process repeats for every word generated by Artificial Intelligence (AI). XCENA, a startup with offices in South Korea and the US, aims to solve this systemic problem. Techcrunch.com reports .
The four-year-old startup has developed a chip that brings computing capabilities as close as possible to DRAM memory. This allows standard data operations to be performed without the costly and energy-intensive "travels" between the processor and memory. If implemented on a large scale, this technology could drastically reduce AI infrastructure costs. Consequently, investors are showing great interest: XCENA recently raised $135 million in a Series B funding round, bringing the company's valuation to $570 million.
XCENA founders Jin Kim, Dohun Kim, and Harry Juhyun Kim are experts who previously worked at memory giants like Samsung and SK Hynix. "Over the past decades, CPUs and GPUs have become smarter, but memory has not. We want to change that," says Jin Kim. He emphasizes that the main bottleneck in AI is no longer just computing power, but the efficiency of the memory system.
The company's MX1 chip connects to the processor via CXL (Compute Express Link) technology and processes data directly within the memory module. According to startup representatives, tasks that previously required 10 servers can now be performed on a single server. This is a highly attractive solution, especially for hyperscalers spending billions of dollars on AI infrastructure.
Currently, the MX1 chip is in the prototype stage. According to the plan, mass production of these chips will be launched at Samsung factories by the end of 2026. The company aims to generate its first revenue starting in 2027.
Read “Zamin” on Telegram!
