I think this is already done to some extent (if you put Arc with an intel iGPU they can share work)
#0
hot marsh
vivid kettle
Bruv
umbral tinsel
... General compute performance in Graphics specific api is going to be lackluster at best and crippling slow at worst.
Take a look at the media engine as a really good proxy.
HEVC encode of 720p footage on a 13900k unconstrained 250W+ can do around 180-200FPS depending on bitrate.
QSV-HEVC can hit 350+ on the igpu on even something like a UHD770...
This is on an IGPU that struggles to maintain 35-40W consumption.
Same workload on a 3090 for example using NVENC hits around 750fps at 140W..
Fixed function capabilities are always going to outperform general function, and most workloads are not threaded to optimize around multiple architectures.