The mysterious factor behind the semiconductor product that is about to "surpass" NVIDIA's AI chip

The global semiconductor-chip market is getting more exciting, as a chip designer who used to work for Intel, AMD and Tesla is working to lower the price of AI applications by designing more efficient chips than the giant NVIDIA, according to Nikkei.

NVIDIA is the world's leading manufacturer of GPU graphics chips. (Source: VCG)

"Legendary" chip?

Jim Keller, CEO of Tenstorrent - an American AI chip design startup, is also known as a globally renowned chip designer as the main designer for AMD's Zen graphics chip line, helping the company "revive" after a long period of struggling to keep up with Intel in the late 2010s.

He is a pioneer in developing the operating system of a set of chips (chipset) for Autopilot - the self-driving software of electric car company Tesla and is called "a legend" by the global chip design community.

Keller told the media that there are many markets that NVIDIA is not serving well. He said that with more AI applications being integrated into high-tech devices such as smartphones, electric vehicles and cloud storage services, companies are looking for cheaper solutions and there are “a lot of small companies that don’t want to pay $20,000” for a high-end graphics processing unit (GPU) from NVIDIA, which is currently considered the best option on the market.

“New” and “different” approaches

Keller’s Tenstorrent, founded in 2016, is preparing to start shipping its second-generation general-purpose AI chips later this year. The company says that in some areas, Tenstorrent’s products outperform NVIDIA’s AI GPUs in power and processing power. The company’s Galaxy system is three times more efficient and 33 percent cheaper than NVIDIA’s popular AI server system, DGX.

Tenstorrent is currently designing its products to be “as cost-effective as possible.” However, CEO Keller admits that it will take many years to break the structure of the current “giant” chip industry, which is dominated by a few “players” like NVIDIA.

Keller revealed that Tenstorrent aims to build a technology that can be used across a wide range of products. In a typical AI chipset, the GPU sends data to memory each time a process is performed. This requires the high-speed data transfer capabilities of high-bandwidth memory (HBM), a key component for synthetic AI chips and one that has played a key role in the success of NVIDIA products. However, HBM is often power-hungry and expensive.

To overcome this weakness, Tenstorrent has designed chips with a focus on reducing energy and lowering product costs. Mr. Keller asserted that with this new approach, his company's chip design can replace both GPUs and HBM in some areas of development with AI applications.

Source: https://baoquocte.vn/nhan-to-bi-an-phia-sau-san-pham-ban-dan-chuan-bi-vuot-mat-chip-ai-cua-nvidia-278780.html