Even well-optimized models only likely to use 35 to 45% of compute the silicon can deliver GPU accelerators used in AI processing are costly items, so making sure you get the best usage out of them ought to be a priority, yet the industry lacks an effective way of measuring this, says the Uptime Institute....