Etched's custom Sohu processor offers an innovative approach to AI chip design, delivering tens of times the performance at a lower cost than general-purpose GPUs by focusing on specialization to work with transformative AI models.
Soohu chip
Unique solution for transformers
Etched startup introduced a specialized Sohu processor optimized for working with transformable artificial intelligence models such as Llama 70B. This chip is not designed to universally handle different types of AI models, but it demonstrates unprecedented performance when working with Transformers, outperforming even the latest GPUs from NVIDIA and AMD by tens of times.
Incredible processing speed
Thanks to an innovative architecture specialized for transformer models, the Sohu chip is capable of process up to 500 thousand tokens per second. The server, equipped with eight Sohu processors, is equivalent in performance to 160 modern NVIDIA H100 GPUs. According to Etched, Sohu is 10 times faster and cheaper than NVIDIA's recently announced Blackwell (B200) chips.
Business case for specialization
Etched makes the case for switching to custom chips like Sohu significant costs associated with training and operating modern artificial intelligence models. Given billions of dollars in investment in this area, even a modest 1% improvement in efficiency could justify creating custom chips of our own that are many times better than general-purpose GPUs.
Glossary
- Etched is an innovative company that has developed a specialized Sohu processor for transformable AI models.
- Sohu is the first chip optimized exclusively for transformer AI models, delivering unprecedented performance and cost efficiency.
- Transformers are an advanced type of neural network architecture that underpins modern AI language models such as ChatGPT and GPT-3.
- Llama 70B - Large AI language model with 70 billion parameters used to demonstrate the capabilities of the Sohu processor.
Links
Questions answered
What is Etched and what is their approach to building AI processors?
How powerful are Etched's Sohu processors compared to NVIDIA chips?
Why does Etched think dedicated AI chips are better than general purpose GPUs?
Why did Etched focus on transformable models?
What are Etched's future plans for custom AI chips?
Hashtags
Save a link to this article
Discussion of the topic – Etched Announces Revolutionary AI Chip: More Powerful Than 160 NVIDIA H100
Startup Etched has introduced an innovative approach to creating AI chips based on the ASIC principle. Their flagship Sohu chip promises 160 times the performance of the NVIDIA H100, specializing in transformative AI models.
Latest comments
8 comments
Write a comment
Your email address will not be published. Required fields are checked *
Hélène
Incredibly interesting news! I completely agree that dedicated processors like Etched's Sohu could be revolutionary for Transformers and generative AI models. 🤖 Agree, performance is 10 times higher at lower costs than NVIDIA competitors - this is simply explosive potential! 💥
Григорий
Yes, progress does not stand still! 🔥 To be honest, the glaring truth is that general purpose GPUs have not progressed that much in recent years. Maybe specialization really is the way forward for the most demanding tasks like huge language models.
Ян
It’s kind of scary to think that in a couple of years every big AI model will work on such specialized chips... 😰 Maybe this is an overly radical view of the Sohu authors? Still, the flexibility of universal solutions is always at a premium.
Марио
And now I’m wondering how difficult and expensive it will be to train such specialized models as in Sohu? 🤔 I doubt that this is available to the mass user... But large corporations and cloud services will definitely benefit by transferring everything to custom solutions like these chips.
Збигнев
Everything has its pros and cons, it is important to find a balance. 💯 On the one hand, specialized solutions like Sohu promise high efficiency for specific tasks. But on the other hand, versatility allows you to flexibly apply technologies in different areas. In my opinion, the future lies in a symbiosis of universal and specialized chips/systems. 🧠
Рудольф
Feh, we found something to argue about! 🙄 All these fancy things like chats with AI are just pure nonsense. In my time, we assembled computers ourselves from radio components, that’s where progress was made! And today's dunces can only wag their tongues... 😒
Анна
Oh, Rudolph, don't be such a grump! 😄 Don’t you notice how much our lives have changed thanks to new technologies? Yes, there are excesses and unnecessary noise around trends. But shouldn't we be excited about real breakthroughs like these custom chips? 💡 For me, this opens up completely new horizons for AI!
Петр
I agree that it sounds very promising. But we should not forget about the potential risks and negative consequences of such powerful and highly specialized technologies 🚧. Ethical issues, safety, potential for misuse - there is a lot to think about here. After all, progress is wonderful, but only if done responsibly.