AMD Acquires AI Chip Startup Taalas to Boost Inference Performance
AMD has acquired Taalas, a startup that specializes in developing custom silicon by hardwiring specific AI models directly onto transistors. This acquisition aims to significantly boost AI inference performance, with early demos showing speeds of up to 17,000 tokens per second. As the AI industry shifts focus toward cost-effective inference, model-specific silicon offers a highly efficient alternative to general-purpose GPUs. This acquisition strengthens AMD's position against Nvidia by providing specialized hardware that can drastically reduce the cost of running large language models. Taalas' architecture unifies storage and compute on a single chip at DRAM-level density, eliminating traditional memory bandwidth bottlenecks. While highly efficient, these chips are hard-wired for one specific model, meaning they lack the flexibility of general-purpose processors.
## BACKGROUND
AI inference is the process of running trained machine learning models to make predictions or generate text, which is computationally expensive on general-purpose GPUs. Taalas addresses this by etching the weights of a specific AI model directly into the chip's hardware circuitry. This application-specific integrated circuit (ASIC) approach bypasses the need to constantly load model weights from external memory, drastically reducing power consumption and latency.