[why_only_15]

Compute shaders are not capable of using modern GPU features like tensor cores or many of the other features needed to feed tensor cores data fast enough (e.g. TMA/cp.async.shared)