"frankly a little shortsighted"
But if you do need GPU co-processing support in a server wouldn't it be better it go for a full-feature dedicated GPU (or two, or however many you can link together these days), rather than a cut-down version built into the CPU die?