Hello everyone
I’d like to run something past you …
Over the past few months I have been trying to figure out what the best options are other than GPUs as I am convinced that GPUs are not the way forward, for the following reasons.
- GPU’s are non deterministic because of a multitude of reasons one of them is operating temperature (thermal throttling) a slight variance from one GPU to the next and it won’t work.
2)All GPU’s need to be the same brand and model this carries over into the manufacturing of GPU’s a tiny deviation in manufacturing and it’s non deterministic - even the same brand same product same factory.
3)Energy and associated running costs are high.
Energy will become more expensive before it gets
cheaper. I heard Dominic Williams say that you might be considering running the GPUs at 60% this makes it costly.
- it’s old technology we are fortunate to have not already commited to GPU’s on the network and this will be of huge advantage. Chatgpt Grok etc have spend billions on GPUs and are now realising that purpose built chips are the way forward (Elon is about to buy Intel)
With the emergence of specialized chips like Tensor Processing Units (TPUs) and Deep X chips. These newer chips are specifically designed for machine learning and artificial intelligence tasks, offering superior performance and energy efficiency compared to traditional GPUs.
-TPUs (Tensor Processing Units): Developed by Google, TPUs are optimized for Google’s machine learning framework, TensorFlow. They excel at tasks like deep learning, natural language processing, and image recognition.
- Deep X Chips: These are a broad category of AI-specific processors designed by various companies, including Cerebras Systems and Graphcore. They often feature massive parallel processing capabilities and specialized memory architectures to accelerate AI workloads.
By focusing on these specialized chips, the network can leverage the most advanced hardware for its AI and machine learning needs, potentially leading to: - Increased performance: Specialized chips can significantly outperform GPUs in AI tasks, leading to faster training times and improved model accuracy.
In conclusion, while GPUs have been a valuable tool for many years, the advent of specialized chips like TPUs and Deep X chips offers a compelling alternative.
Surely it’s a case of rent GPU’s for the interum?
Another avenue for thought is AI on edge in the not to distant future (3 years - or less) we will see AI chips on our phones giving everyone the power to run ai on edge.
I can’t reiterate this point enough but think about specifying GPU’s over tpus and AI on edge.
Time is in our side here.
You have an advantage in that the road is open make the smart choice and break away form the heard.
Kind regards
Kurt