Recently, Etched, a company in Silicon Valley, is garnering widespread attention. This startup, founded by two Harvard dropouts, has introduced a chip named Sohu, which they claim is 20 times faster than Nvidia's current top-of-the-line H100 chip when running AI models like ChatGPT. This news has sparked curiosity about what Sohu is and how it could have such a significant impact on the AI hardware market.
Image source: Image generated by AI, provided by Midjourney
Sohu's design philosophy is quite unique. Unlike Nvidia's GPUs, Sohu uses an Application-Specific Integrated Circuit (ASIC) chip, which is specifically optimized for Transformer AI models. This means that Sohu cannot run other AI models such as Convolutional Neural Networks (CNN), but it offers superior computational power for tasks like ChatGPT, OpenAI's Sora, and Google's Gemini. This specialized design could enable developers to explore new products that were previously unattainable, such as real-time translation applications that can instantly translate conversations between multiple languages.
When looking ahead to Sohu's potential applications, we can't help but think about its capabilities in multimodal processing. Sohu can not only handle text but also integrate images and language, opening up new possibilities for visual question answering and other interactive applications. However, this is still theoretical, as Etched has successfully raised $120 million to drive the development of Sohu, but the actual release date remains uncertain.
It's worth noting that Etched has already received tens of millions of dollars in pre-orders for its hardware and has entered into a partnership with TSMC, the world's leading semiconductor manufacturer, to produce this 4nm chip. For Etched, this deal will significantly boost their production capabilities and help them succeed in the market.
Sohu chip from Etched could bring a revolution to the AI industry, but we still need to be patient and see if this chip can deliver on its promises.
Key Points:
🌟 Sohu chip claims to be 20 times faster than Nvidia's H100 when running AI models.
⚙️ The chip is designed specifically for Transformer AI models and cannot handle other AI models.
💰 Etched has raised $120 million and partnered with TSMC for production.