Taalas’ technology and performance claims
Taalas focuses on model-specific integrated circuits, a new approach that embeds the weights and dataflow of an AI model directly into the silicon. This reduces the need to move data in and out of high-bandwidth memory, a common bottleneck. The startup has already produced a test chip, called HC1, which was made using TSMC’s 6-nanometer manufacturing process. HC1 demonstrated performance on Meta’s Llama 3.1 8B model at nearly 17,000 tokens per second. According to Taalas, this is 73 times faster than Nvidia’s H200 chip while using just 10% of the power.
The main tradeoff with this design is that each chip is built for one specific model and cannot be reused for others. However, Taalas argues that this limitation is more manageable than it might seem. The company says that when updating a model, only two out of over 100 layers in the chip design change. Additionally, using in-house tools allows for a tape-out time of about two months between different versions. This efficiency helps reduce costs and development time for new models.
AMD’s AI roadmap and integration plans
AMD plans to integrate Taalas’ chips into its broader AI ecosystem. The move will see Taalas hardware paired with AMD’s Instinct GPUs, Helios racks, and Epyc processors, all managed under the ROCm software stack. This strategic combination offers a more flexible and powerful platform for customers. Vamsi Boppana, AMD’s senior vice president in the AI Group, highlighted the value of this acquisition, stating it provides customers with a wider range of compute options tailored for various AI workloads.
Boppana emphasized the performance and efficiency gains Taalas brings, calling them 'differentiated.' Taalas CEO Ljubisa Bajic, who previously led Tenstorrent, a now-defunct AI chip startup, founded Taalas with the goal of completely rethinking how inference is done. He described the company’s mission as building hardware that aligns precisely with the model, a shift from traditional, flexible but often inefficient approaches.
Taalas’ funding and AMD’s AI strategy
Taalas has been backed by major investors, including Quiet Capital, Fidelity, and semiconductor investor Pierre Lamond. The latest funding round in February added $169 million to the startup’s coffers, bringing its total capital to around $219 million. AMD’s acquisition of Taalas is part of a broader AI expansion push. Over the past nine months, the chipmaker has acquired MK1 in November, Mext in June, and integrated FastFlowLM into its operations in July. Each of these deals has helped AMD strengthen its position in the AI hardware space.
The growing interest in inference is reshaping the AI hardware market. For example, Nvidia entered into a $20 billion licensing deal with Groq last December, and Groq used that technology to launch the Groq 3 language processing unit at its GTC event in March. AMD’s acquisition of Taalas signals its intent to compete in this fast-moving segment. By bringing a high-efficiency inference solution into its portfolio, AMD is aligning itself with the market’s shift toward specialized silicon optimized for AI tasks.