Sohu (AI processor)

2024: Product Announcement

On June 25, 2024, startup Etched announced a super-powerful accelerator called Sohu for artificial intelligence tasks. It is said to be tens of times faster than Nvidia's GPU-based solutions.

Etched was founded in 2022 by Harvard graduates Gavin Uberti and Chris Zhu, who, together with Robert Wachen and former Cypress Semiconductor CTO Mark Ross, sought to create a chip capable of working with AI models as efficiently as possible. The developed Sohu product is a special purpose integrated circuit (ASIC): it is designed to work exclusively with transformer models. This architecture, proposed by Google researchers in 2017, is dominant in the generative AI market as of 2024.

𝓲

Etched

The Sohu processor is manufactured using 4-nanometer TSMC technology. It is claimed that the novelty is capable of providing significantly higher output performance (up to 20 times) than Nvidia GPU accelerators and other general-purpose AI chips, while consuming less power.

Sohu is an order of magnitude faster and cheaper than even the next generation Nvidia Blackwell accelerators GB200 when working with text, image and video converters. One Sohu server can replace 160 Nvidia H100 GPUs. Sohu will be a more affordable and effective option for teams that need specialized AI solutions, says Uberti.

It is noted that Sohu uses a hardware-software output pipeline optimized for transformers. This made it possible to abandon the components that are required to support other AI models and reduce software overhead.^[1]

Notes