Nvidia’s H100 and B200 Graphics Processing Units (GPUs) are the reason you can ask ChatGPT questions like “How to eat a mango without spilling it” and many more. With a market valuation that surpassed both Apple and Microsoft by $3 trillion last month, Nvidia has emerged as the leader in the AI hardware sector thanks to these remarkable chipsets that enable AI chatbots.
However, a relatively new company led by two Harvard graduates is now vying for market dominance in AI hardware. With its transformer ASIC (Application Specific Integrated Circuits) chip, named Sohu, Etched, a California-based firm, hopes to upend the AI chipset market.
Also Read: Google Set to Acquire Cybersecurity Firm Wiz for $23 Billion
When it comes to operating transformers like ChatGPT, Sohu promises to be 20 times faster than Nvidia’s flagship H100 GPU. Based on emulation studies, the company claims that the B200, the more powerful Nvidia offering than the H100, is reportedly 10 times slower than the Sohu.
To provide high computing power to run billions of parameters (variables that are utilized in training an AI model) for transformer models, Sohu is adopting a completely different strategy. Instead of creating a specialized chip that supports only transformer AI models—those that run ChatGPT, Sora (OpenAI’s text-to-video AI model), and Google’s Gemini—Etched is opting to design a specialized semiconductor that caters to several computationally demanding activities, such as generating visuals in real-time.
Also Read: Google Gemini AI Reads Private Drive Docs Without User Consent: Twitter Revelation
This implies that other AI models, such as Convolutional Neural Networks, which are used for image recognition, cannot be run on it. This allows developers to investigate new AI products that were previously unfeasible because of GPU power limitations.
For instance, Sohu may eventually result in a real-time translator that can converse in French, English, and German while hearing and reading in Hindi, Gujarati, or Tamil. However, in theory, it opens up the prospect of multimodal and multilingual translation, which of course requires more than simply processing capacity.
The chipset may also use transformers to integrate the visual and verbal domains, which is another multimodal use. This means that a model of that kind would be able to comprehend text and visuals at the same time, opening the door to the potential of visual questions and replies, much as in an interview.
Also Read: Perplexity AI is releasing its Pro Search function for intricate queries on Android Devices
However, this is all still just speculation. To make it happen, Etched raised USD 120 million on June 25. However, a precise date for the Sohu ASIC’s release is yet unknown.
According to Etched, preorders for the hardware, already totalling “tens of millions of dollars” have been placed. To manufacture the 4-nanometer chip, the business has also negotiated a contract with Taiwan Semiconductor Manufacturing (TSMC), claiming that the agreement will enable it to “ramp up our first year of production.”
Also Read: SpreadsheetLLM by Microsoft: A Game Changer for Data Analysts and Accountants
