Are GPU days numbered? Feels like we just got started right? The playing field’s moved to quantum algorithms that can handle really complex datasets now, not just memorizing data lakes.
Microsoft is at the forefront of retail quantum AI with Tensorflow Quantum supported on Microsoft’s Majorana 2.
This is nowhere near a death blow to Nvidia, for now. GPUs will still be needed to augment agentic tasks because Tensorflow builds classic machine learning models, specifically hybrid-quantum classic models.
That means it combines classic neural networks with quantum computing capabilities and that comes with limitations:
- The quantum components must be precise and performance is heavily dependent on where integrated circuits are placed within the model.
- It’s really hard to make changes, they must be big to see any visible performance gains. A complete model overhaul may be required.
- Most important is some quantum resources, such as qubit volume and circuit depth can be limited.
In plain terms, it’s a pain. Microsoft seems to have the architecture sorted out to overcome limitation #1. Limitation #2 reminds me of having to install DOS six times in 1994 to get the BIOS to recognize a new ISA card. If it works it works, if it doesn’t take the machine apart and start over. #3 is hard to overcome, purely resource dependent, and where money becomes the fastest path to pseudo-sentient intelligence.
The question I keep asking, and haven’t received a clear answer on, is why are we still investing in and building enormous data centers for GPU based algorithms when the near future is hybrid-quantum circuits and closed-weight quantum models? By putting it that way I mean to imply the benefit of doubt to our esteemed AI model Gods, may they go forward without an open source quantum model capable of self-inference post training on any resources of scale. If “they” haven’t done it already.
