I mean, are they? Seems like the industry would prefer these things to become commodities, especially if it helps with portability and reproducibility.
The specialized chips may very well become commoditized over time. It is the way things go if the market is large and competitive enough - and that there is sufficient competitive advantage over the existing types of devices (GPUs in this case). For a rather recent example, see Bitcoin mining ASICs.
Of course GPUs (and CPUs, and pre-transformer/LLM type NPU/TPUs) are also going to respond to better suit LLM workloads. So we may see a convergence in capabilities.
What is for sure is that the LLM chip architecture is not yet settled. Currently available chips are mostly designed pre LLM craze, with slight adaptations. And have large areas of potential improvement, at least 10-100x (maybe 1000x) in power efficiency. And power efficiency is a key element (in addition to up from hardware investment). Which is the opportunity this startup (and others) have identified.
Given that Moore’s law is slowing down, given that circuit design is getting more automated and given the rise of chiplet architecture, I expect that you will have more specialized chips.
Of course GPUs (and CPUs, and pre-transformer/LLM type NPU/TPUs) are also going to respond to better suit LLM workloads. So we may see a convergence in capabilities.
What is for sure is that the LLM chip architecture is not yet settled. Currently available chips are mostly designed pre LLM craze, with slight adaptations. And have large areas of potential improvement, at least 10-100x (maybe 1000x) in power efficiency. And power efficiency is a key element (in addition to up from hardware investment). Which is the opportunity this startup (and others) have identified.