AMD launches small language model focused on optimizing tasks through artificial intelligence

AMD, which has been focusing on the hardware market for a long time, changed its direction and announced its first small language model (SLM). The technology, called AMD-135M, was created for enterprise use to optimize specific tasks.

AMD-135M uses a technique called speculative decoding to perform task optimization. This technique makes the entire process more logical by making informed predictions about future coin demands when creating existing tokens in the processing pipeline.

According to the company, 135M was trained from scratch with 670 billion data tokens. This process took approximately six days using four Instinct MI250 AI accelerators, and AMD further enhanced the model with 20 billion tokens focused on coding.

posture change

It’s worth repeating that AMD’s first SLM had variants: AMD-Llama-135M and AMD-Llama-135M-Code. As the name suggests, these language models are based on the Llama family and were created to meet the needs of the company’s customers who need new pre-trained models.

The release of the model also shows that AMD’s stance has completely changed. Recently, the company led by Lisa Su confirmed that it will focus its efforts on developing solutions for artificial intelligence and less on gaming graphics cards.

This doesn’t mean the red team will end GPU production, but AMD tends to become more of a rival to Nvidia when it comes to software and other technologies.