Startup Etched raises US$120m to bet on transformer-only ASIC

Startup Etched raises US$120m to bet on transformer-only ASIC

Business news |
By Peter Clarke Inc. (Cupertino, Calif.), a 2022 startup, has raised US$120 million in Series A round of funding to develop a transformer-only ASIC called Sohu and take on Nvidia.

The company claims that because Sohu is architected to run transformer models, which are at the heart of LLMs and such applications as ChatGPT, it is 20x faster than Nvidia H100s GPUs.

The downside of that is that Sohu cannot execute many traditional AI models that a GPU can. Sohu can’t run CNNs, RNNs or long short-term memory (LSTM) models, deep learning recommendation models (DLRMs); protein-folding models, such as AlphaFold 2, or older image models such as Stable Diffusion 2, the company states.

But Etched is fine with that because it reckons almost every state-of-the-art AI model these days is based on transformer software. These include ChatGPT, Sora, Google’s Gemini, Stable Diffusion 3, and many others. “If transformers are replaced by SSMs [State Space Models], RWKV [Receptance Weighted Key Value], or any new architecture, our chips will be useless,” the company admits on its website.

Or as CEO Gavin Uberti put it when speaking to CNBC: “If transformers go away, we’ll die. But if they stick around, we’re the biggest company of all time.”

When models cost more than US$1 billion to train, specialized chips become inevitable Etched argues. In a multibillion dollar annual market a specialized chip can be justified for a small percentage improvement in performance. But as it stands transformer ASICs are orders of magnitude faster than GPUs at executing transformer software. Just as specialized bitcoin mining chips replaced GPUs back in 2014 so will GPUs be replaced by specialized chips in AI, Etched claims.

Sohu only supports transformer inference, whether Llama and Stable Diffusion 3. But because Sohu is optimized for one algorithm, the vast majority of control flow logic can be removed, allowing it to have many more math blocks.  Sohu also boasts over 90 percent FLOPS utilization, compared with about 30 percent on GPUs running transformer software, Etched said.

According to reports the company was founded by CEO Gavin Uberti and Chris Zhu, a teaching fellow at Harvard University.

It will use the US$120 million raised to tape-out its ASIC and get it manufactured by TSMC. No details have been provided of the manufacturing process node being targeted or when chips may be available for sampling.

Deep learning transformer software is so called because of a software model’s ability to process input data while understanding the context and relationships within the data, according to an Nvidia blog. The transformer model uses a mechanism called ‘attention’ or ‘self-attention’ that allows it to focus on different parts of the input data and understand how each part influences the others. This is useful in processing sequential data such as text, where the meaning of a word can depend on the words around it.

The selection of the term transformer may also reflect the film franchise of the same name that was popular at the time the term was coined by Google researchers in 2017.

Related links and articles:

News articles:

Nvidia continues to quadruple sales year-on-year

Nvidia pushes Blackwell GPU as start of a whole new industry 

AMD announces AI roadmap through 2026   

Intel aims at Nvidia with Gaudi 3 AI chip  

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles