Interesting move by AMD:
https://www.servethehome.com/amd-to-acquire-taalas-for-model-specific-ai-inference-chips/
Basically yet another iteration of the "if software is too slow, make it hardware" continuum. This time a AMD acquisition turns LLMs into single-model-chips.
Curious to see if and how this will play out to commercial usability and on what type of model.
Also:
- Might we even see coprocessor boards with exchangeable model chips?
- What will the energy consumptions of those look like compared to the full generals purpose inference runner setups?
#amd #hardware #servethehome #llm #customsilicon #ai #asic #innovation #mergersandacquisitions