[Chipformation] Samsung Tailor's First Client, AI Chip Startup 'Groke'

Fabless Company Founded by Former Google Engineers
Taylor to Produce 4nm AI Accelerator Semiconductor Chips

Samsung Electronics' first customer for its semiconductor foundry plant under construction in Taylor, Texas has been revealed. It is the AI chip startup Groq.

Groq is a US semiconductor design (fabless) company founded in 2016 by former Google engineers, including Jonathan Ross, who was a chip engineer for Google's machine learning. It is headquartered in Mountain View, founded by Fairchild, the pioneer of Silicon Valley companies.

Groq, a US semiconductor startup, logo.
[Photo by Groq website]

Groq integrates eight of its PCIe (a serial interface standard supporting high-speed data input/output) AI chips into a single chassis and ships the TSP (Tensor Streaming Processor) for AI inference in data centers to the server market. According to figures released by Groq, the TSP boasts powerful performance of 1 POPS (1,000 times 1 TOPS, performing 1,000 trillion operations per second). In terms of batch size (the number of data processed at once), it recorded 18,900 IPS (inferences per second) on ResNet-50 v2.

Accordingly, Groq's next-generation AI chip is known to have up to four times higher power efficiency and superior performance compared to existing products. Compared to complex existing architectures based on CPU (Central Processing Unit), GPU (Graphics Processing Unit), and FPGA (Field-Programmable Gate Array), Groq's chip offers the advantage of simplifying adoption and application. Therefore, it enables customers to quickly and easily implement high-quality systems that require scalability and high performance per watt.

Groq is expected to produce 4-nanometer (nm, one billionth of a meter) AI accelerator semiconductor chips at Samsung Electronics' Taylor plant. These chips are expected to increase the interconnection bandwidth of the Language Processing Unit (LPU) system, delivering performance superior to current exascale supercomputers. Groq plans to configure the chips in quantities ranging from 85,000 to over 600,000 to build customized enterprise systems.

Jonathan Ross, CEO and founder of Groq, said, "Our first-generation LPU was designed from the ground up for AI, enabling consistently superior performance over GPUs in AI and ML (machine learning) tasks," adding, "Through our partnership with Samsung, we will continue to leverage the most advanced semiconductor manufacturing technology to make these leaps."

This content was produced with the assistance of AI translation services.