## AMD boosts AI inference with Taalas acquisition

AMD boosts AI inference with Taalas acquisition

AMD has bought the AI‑chip startup Taalas. The deal lets AMD put neural‑network models directly into silicon. By etching the model weights onto the chip, the hardware can run inference without loading large files from memory. This reduces latency and power use. It also simplifies deployment for edge devices that need fast AI responses. The move positions AMD to compete with other vendors that offer fixed‑function AI accelerators.

The new approach works like a tiny, dedicated processor for each model. Once the model is programmed into the chip, the hardware executes the calculations in a single pass. A simple illustration of the workflow is shown below:

// Load model weights at manufacturing time (fixed in silicon)
const float weights[NUM_WEIGHTS] = { /* etched into chip */ };

// Run inference on input data
float result = infer(weights, input_tensor);

The inference step runs entirely on‑chip, eliminating memory transfers. AMD says this will improve performance for cloud and edge AI workloads. For more details, see the original article: AMD acquires Taalas to boost inference performance by etching models in silicon.