AMD Completes Acquisition of AI Inference Chip Company Taalas
AMD announced that it has completed its acquisition of Canadian AI inference chip company Taalas and plans to integrate its technology into its accelerator roadmap, developing system-level solutions alongside AMD Instinct GPUs. Taalas seeks to turn AI models into circuitry embedded within chips. Its HC1, released in February this year, is hardware-optimized for Meta's Llama 3.1 8B and, according to the company, can deliver throughput of 17,000 tokens per second for a single user. The architecture trades versatility in adapting to different models for greater inference efficiency.
Why read this
AI score 71
The acquisition brings model-specific hardware into AMD's accelerator roadmap, highlighting chipmakers' trade-off between inference efficiency and versatility.
Topics
Body · AI translation
Chips and AI reported on November 18 that AMD announced yesterday it had completed its acquisition of Canadian AI inference chip company Taalas.
AMD plans to integrate Taalas’ technology into its accelerator roadmap and jointly develop system-level solutions centered on AMD Instinct GPUs, while creating synergies with its broader product portfolio, including AMD EPYC CPUs, AMD ROCm software, and AMD Helios rack-scale solutions.
Founded in 2023 and headquartered in Toronto, Canada, Taalas is a semiconductor company focused on AI inference chips. Unlike conventional GPUs, which retrieve model weights from memory before performing computations, the company seeks to convert AI models directly into hardware circuits inside chips, making the models themselves part of the computing hardware.
Taalas co-founder and CEO Ljubisa Bajic has many years of chip design experience. From 2003 to 2006, he worked in chip design at Canadian GPU design company ATI Technologies. After AMD acquired ATI Technologies in 2006, Bajic joined AMD, where he successively served as an ASIC design manager and an IC design architect.
In 2013, he joined Nvidia as a senior architect. After working at Nvidia for a year, Bajic returned to AMD. In 2016, he founded U.S. AI chip company Tenstorrent. Renowned chip architect Jim Keller succeeded him as the company’s CEO in 2023. That same year, Bajic founded Taalas.
▲Taalas co-founder and CEO Ljubisa Bajic
Taalas’ flagship product, the HC1 Technology Demonstrator, was released in February this year. Manufactured using TSMC’s 6nm process, the chip is specifically optimized in hardware for Meta’s Llama 3.1 8B model.
▲Taalas’ flagship product, the HC1 Technology Demonstrator
According to the company, the HC1 achieves single-user throughput of 17,000 tokens per second on the Llama 3.1 8B model, about 74 times that of Nvidia’s H200 and about 48 times that of the B200.
▲Throughput data for multiple chips on the Llama 3.1 8B model. Source: Taalas
Notably, AMD has recently completed two transactions involving the AI industry chain. Previously, the company announced that it had completed its acquisition of World Labs, founded by Fei-Fei Li.
Taalas’ HC1 chip demonstrates the performance potential of specialized architectures for AI inference, but its high throughput also comes with certain limitations: its performance advantage depends heavily on a specific model, making it difficult to adapt flexibly to different models and computing tasks as general-purpose GPUs can. In other words, Taalas has chosen to trade versatility for exceptional inference efficiency.
AMD’s acquisition of Taalas and integration of its technology into its accelerator product roadmap also indicate that specialized computing architectures are attracting the attention of mainstream chipmakers. As large models continue to evolve and AI use cases become increasingly specialized, market demand for chips is no longer limited to raw computing-power gains but is extending further to inference speed, energy efficiency, deployment costs, and model compatibility.
AI Trends