Etched Unveils Sohu: The World's First ASIC Chip Hardwired for Transformer Models

Etched announces Sohu, a revolutionary ASIC chip hardwired specifically for Transformer architectures, promising massive speedups over NVIDIA graphics cards.

Silicon Valley hardware startup Etched has shocked the semiconductor industry by unveiling Sohu, the world's first application-specific integrated circuit (ASIC) chip hardwired entirely for Transformer architectures. By sacrificing generic computing flexibility to excel at a single job, Sohu promises performance speedups and efficiency gains that outclass industry-standard general-purpose GPUs like the NVIDIA H100 and B200.

#

The Bet on Transformers

General-purpose graphics processors (GPUs) are designed to run any math model, from convolutional neural networks to graphics rendering. Etched's engineering team took a different bet: that the Transformer architecture will remain the foundation of all future artificial intelligence, including LLMs, diffusion models, and code generators.

"By burning the Transformer's attention mechanism directly into the silicon, we eliminate the instruction overhead and latency that slows down GPUs," said Arjun Krishnan, Senior Hardware Architect. "If you run a Transformer on Sohu, it runs at the speed of custom hardware. If the industry shifts away from Transformers, the chip becomes obsolete. We are making a major bet on this architecture, and the preliminary benchmarks indicate it will pay off."

#

Performance Benchmarks and Hardware Specs

Early engineering tests show Sohu achieving jaw-dropping numbers: * Unmatched Throughput: Sohu can process over 500,000 tokens per second per server node, which is roughly 20x faster than comparable GPU setups. * Lower Power Consumption: By stripping out complex instruction decoders and registers needed for general computing, Sohu runs cooler and uses 50% less electricity. * Cost Efficiency: A single Sohu node can replace up to 8 GPUs for transformer inference workloads, dramatically lowering server hosting and cloud infrastructure overheads.

#

A Shift in the AI Compute Market

As hardware costs continue to constrain AI startups, specialized chips like Sohu could democratize access to high-performance computing. Cloud providers are already lining up to test Sohu chips in their datacenters, looking for ways to reduce their reliance on NVIDIA's supply chain and offer cheaper inference rates to enterprise developers.

Compare startup hardware metrics using the new Startup Benchmarking Compare Tool.