The reason ChatGPT allows you to ask “how to eat a mango without dropping it” and more is because of Nvidia’s H100 and B200 graphics processing units (GPUs). These magical chipsets powering AI chatbots have made Nvidia a leader in the AI hardware industry, with its market capitalization hitting the $3 trillion mark – more than Microsoft and Apple combined – last month.
But now, a relatively young startup founded by two Harvard dropouts has set its sights on the AI hardware pie. Etched, a California-based startup, is trying to disrupt the AI chipset market with its transformer ASIC (application specific integrated circuit) chip called Sohu.
Sohu claims that it is 20 times faster than Nvidia’s flagship- the H100 in running transformers like ChatGPT. According to claims made by the company based on emulation tests, the B200, which is Nvidia’s more powerful offering than the H100, is reportedly 10 times slower than Sohu.
Source: X/@Etched
Sohu is taking a completely different approach to providing high computational power to run billions of parameters (variables used in training AI models) for Transformer models. Unlike GPUs that can perform many computationally heavy tasks (such as rendering graphics in real-time), Etched is opting to build a specialized chip that caters only to Transformer AI models – those that run ChatGPT, Sora (OpenAI’s text to video AI model) and Google’s Gemini.
This means that it cannot run other AI models such as convolutional neural networks (used for image recognition). This opens up the possibility for developers to explore new AI products that were not possible until now due to the limited power on GPUs.
For example, Sohu could potentially lead to a real-time translator that can listen and read Hindi, Gujarati or Tamil and respond in French, English and German. Of course, this kind of multimodal and multilingual translation requires more than just computational power, but in theory, it opens up the possibility.
Another multimodal application of Transformers that the chipset can use is to integrate the visual and language fields. This essentially means that such a model will understand both text and images simultaneously, opening up the possibility of visual questions and answers, like in interviews.
But this is all just a theory. On June 25, Etched raised USD 120 million to make this a reality, the actual timeline for the actual release of the Sohu ASIC is still unclear.
Today we announced a $120 million fundraise to bring Sohu to the world. Thank you @edludlow Thanks for hosting us! pic.twitter.com/Q4SeuUjP2h
— Etched (@Etched) June 26, 2024
Etched claims it already has “tens of millions of dollars” worth of hardware reserved in preorder. The company has also secured a deal with TSMC (Taiwan Semiconductor Manufacturing Company) to make 4-nanometer chips, promising that the deal will help “scale up our first year of production.”
