AMD acquires AI startup Taalas to integrate custom inference hardware

AMD has announced the acquisition of Taalas, a startup specializing in custom, non-programmable AI inference chips. Taalas’s technology is designed to bake specific AI models directly into silicon, bypassing the flexibility of traditional GPUs to achieve significantly higher performance. The company’s HC1 technology demonstrator, built on a 6nm TSMC process with 53 billion transistors, reportedly achieves up to 17,000 tokens per second on the Llama 3.1 8B model. This performance level is substantially higher than the throughput of standard hardware like the Nvidia H200 or Groq chips. Taalas’s approach focuses on data flow, where compute is optimized for the specific movement of weights and activations within the chip’s architecture. Following the acquisition, the Taalas team will report to Vamsi Boppana, AMD’s head of AI, and will be integrated into the company’s AI group. The acquisition is expected to close by the end of the year, with the Taalas team likely maintaining a degree of operational independence to continue their work on model-specific inference hardware.

Tech4K StaffVideo report

AMD has announced the acquisition of Taalas, a startup specializing in custom, non-programmable AI inference chips. Taalas’s technology is designed to bake specific AI models directly into silicon, bypassing the flexibility of traditional GPUs to achieve significantly higher performance. The company’s HC1 technology demonstrator, built on a 6nm TSMC process with 53 billion transistors, reportedly achieves up to 17,000 tokens per second on the Llama 3.1 8B model. This performance level is substantially higher than the throughput of standard hardware like the Nvidia H200 or Groq chips. Taalas’s approach focuses on data flow, where compute is optimized for the specific movement of weights and activations within the chip’s architecture. Following the acquisition, the Taalas team will report to Vamsi Boppana, AMD’s head of AI, and will be integrated into the company’s AI group. The acquisition is expected to close by the end of the year, with the Taalas team likely maintaining a degree of operational independence to continue their work on model-specific inference hardware.

AMD has entered into a definitive agreement to acquire the AI startup Taalas. Taalas specializes in non-programmable inference chips that bake specific AI models directly into silicon.

The Taalas HC1 demonstrator chip is built on a 6nm TSMC process and contains 53 billion transistors. Taalas claims its hardware can achieve 17,000 tokens per second on the Llama 3.1 8B model.

The Taalas team will report to Vamsi Boppana, AMD's head of AI, following the acquisition's expected close at the end of the year.

Chapter guide

Introduction to AI AcceleratorsAn overview of what AI accelerators are and how they differ from traditional GPUs.The Taalas ApproachAn explanation of Taalas's unique strategy of baking AI models directly into silicon.HC1 Performance BenchmarksA breakdown of the HC1 technology demonstrator's specifications and performance metrics.AMD Acquisition AnnouncementDetails regarding the acquisition of Taalas by AMD and its implications for the AI market.Expert DiscussionA conversation with Sally Ward-Foxton from EE Times regarding the acquisition and the future of AI hardware.

Worth noting

  • The acquisition is subject to standard regulatory procedures and is expected to close by the end of the year.
  • The performance figures for the HC1 chip are based on a technology demonstrator and may not reflect final production hardware.
  • The specific financial terms of the acquisition were not disclosed.

Watch the original video ↗