[sliken]

Yes it's possible, but at the cost of 2-3x less memory bandwidth, which is a key feature to provide GPU and large language model performance.