Benchmarks

Device Performance

Generally adding just 1 CPU-core to a GPU doesn't add more performance, but decrease it if data doesn't fit cache and both devices share same memory. I wish I had a VEGA or VOLTA GPU to compare but these are all I've got.

Celeron N3060 with 2 cores @1.6GHz and 1 iGPU(HD-400) with 12 cores @600MHz (laptop on battery):

	CPU-C#	CPU-1core-OpenCL	iGPU OpenCL	CPU+iGPU OpenCL
c=sqrt(a)+pow(b,0.5) for 1M elements, 100 repeats	18118ms	22409ms	3211ms	4450ms
c=exp(sin(a)+cos(b))	3523ms	1196ms	174ms	292ms

FX8150 with R7-240(320 cores):

	CPU-C#	CPU-7core-OpenCL	GPU OpenCL	CPU+GPU OpenCL
c=a+b for 4M elements	39.9ms	4.7ms	12.4ms	4.5ms
image resize from 2592x1944 to 512x512	23.5ms(System.DrawImage)	6.4ms	6.3ms	5.7ms

• pipelining increases performance when compute latency is comparable to buffer access/copy latency and only if there are enough cpu threads to control all work-scheduling fast enough. Otherwise, it decreases performance as in this "low-end laptop on battery" example.