Is there any sense in which integrated GPUs outperform dedicated GPUs when matched spec for spec?
In other words, does the GPU being embedded on the same chip as the CPU provide any advantages in terms of bandwidth/latency that means they can outperform matched GPUs which have to communicate via PCIe?
Latency may be slightly improved but bandwidth takes a massive nosedive. GPUs need memory too. Sharing memory with the CPU is far from ideal because doing so not only takes RAM resources from the CPU, it must also share the available bandwidth. To top it off, DDR4 RAM is slow compared to GDDR5, GDDR6, and HBM.
What about in the future?
Even if 3D stacked HBM RAM is placed on the CPU die exclusively for the integrated GPU, the CPU and GPU will be competing for "thermal resources". That is to say, when the CPU heats up, so does the GPU and vice-versa. In applications and games that make use of both simultaneously, the CPU, GPU, or both may need to throttle their clockspeeds to avoid overheating. This obviously degrades performance.