AMD revealed its first simplified language model for artificial intelligence, the AMD-135M – belonging to the Llama line and aimed directly at the corporate market.
This is the latest initiative from the manufacturer, which not only intends to bring new hardware to enable the intensive work of AIs, but is also working on software development – trying to reach a segment in which NVIDIA already operates.
See also:
AMD’s new release will come in two versions: AMD-Llama-135M and AMD-Llama-135-code – each designed to optimize specific tasks by accelerating inference performance using speculative decoding technology. Both are being tested by the manufacturer and there is no scheduled date for their arrival.
Check out the definitions provided by AMD for each model below:
- The base model, AMD-Llama-135M, was trained using 670 billion general data tokens. The process took six days using four AMD Instinct MI250;
- The AMD-Llama-135M-code was refined with over 20 billion tokens especially focused on coding, completing its task in four days using the same hardware
AMD and speculative decoding
One of the main reasons why the AMD-135M and the encode version are so fast is because of speculative decoding. With it, a small “scratch model” generates multiple tokens and a single pass. They are sent to a larger model, which checks and corrects incorrect information.
With this action, multiple tokens are generated simultaneously and allow greater agility during the process. However, this does not come without an energy cost – with a large increase occurring through increased data transit.
AMD believes that with optimizations, the AMD-135M can bring even better performance in the future. Considering they are presenting data with the Instinct MI250, it’s worth noting that the MI300X is already on the market and the next generation Instinct MI325X GPUs are in the final stages of development.
It was not revealed whether this new simplified model of artificial intelligence had its production helped by AMD’s recent acquisitions – such as Silo AI and ZT Systems. However, this seems to be the path they want to take from now on, meeting market needs with AI models pre-trained on their hardware.
Source: Tom’s Hardware
Join the Adrenaline offer group
Check out the main offers on hardware, components and other electronics that we found online. Video card, motherboard, RAM memory and everything you need to build your PC. By joining our group, you receive daily promotions and have early access to discount coupons.
Join the group and take advantage of promotions
Source: https://www.adrenaline.com.br/amd/amd-135m-amd-revela-primeiro-modelo-de-linguagem-simplificada-para-ia/
