Taalas Inc., a startup specializing in chips optimized for specific artificial intelligence models, has secured $169 million in new funding. This significant investment, reported by Reuters, includes backing from prominent investors like Quiet Capital and Fidelity. The latest round pushes Taalas' total funding past the $200 million mark.
The company's initial product is a chip engineered to execute the open-source Llama 3.1 8B language model. Taalas claims its chip delivers 17,000 output tokens per second, outperforming Nvidia's H200 graphics card by 73 times while consuming one-tenth of the power.
This enhanced efficiency stems from tailoring chips to individual AI models, eliminating redundant components. Unlike general-purpose graphics cards, Taalas' custom design swaps unused RAM for transistors that accelerate model performance.
Taalas engineers have reduced development costs by customizing only two of the over 100 layers within their chips. These custom layers incorporate a mask ROM recall fabric, where each module stores four bits processed by a single transistor for matrix multiplications, the core calculations for AI decision-making.
This design eliminates the need for High Bandwidth Memory (HBM) modules, which can slow down processing due to data transfer delays. Taalas' approach bypasses these delays and the associated support components.
The company is reportedly developing a new chip for a Llama model with 20 billion parameters, slated for release this summer. A more advanced processor, HC2, capable of running frontier models, is also planned.