China is looking into new ways to rely less on CUDA software of Nvidia, which is important for the company’s leadership in AI.
Wei Shaojun, who is an executive at the China Semiconductor Industry Association, told the AI industry of China to come up with alternatives to Western technologies like CUDA.
CEO of Nvidia, Jensen Huang has over and over mentioned CUDA as the greatest advantage of the company, highlighting the importance of its software ecosystem in promoting adoption.
Developers reguraly use the CUDA, which also connects them to the hardware of Nvidia. They use it because it is well-established and supported.
Wei Shaojun suggested a different approach which is called software-defined chips (SDCs) rather than making a direct replacement for CUDA.Instead of depending on set hardware designs, this notion transfers more of the processing logic into software.
Developers can run workloads in an SDC system without a CUDA-like layer. chips use a grid that can be changed by using instructions generated by the compiler.
This increases the flaxibilityof the system since the software as well as code are not dependent on a specific set of hardware instructions.
SDCs use deterministic compilation, which means that data transfer is preplaned and controlled down to exact time, in contrast to conventional GPUs that depend on schedulers to handle workloads.
Wei Shaojun pointed out that it would take a lot of resources to create CUDA alternatives using translation layers and distinct ecosystems.
Despite its difficulties, he said that the software-defined chip approach is a more realistic choice. These include intricate compiler specifications, branching and routing problems, and design modifications that deviate from conventional hardware models.
Similar ideas underpin several current solutions, such as those created by businesses like Groq and SambaNova Systems.
These technologies, however, are not direct substitutes for GPUs and are usually built for certain workloads.
